@mutagent/cli 0.1.116 → 0.1.117

package/dist/bin/cli.js

@@ -1241,9 +1241,9 @@ var init_sdk_client = __esm(() => {
 
 // src/bin/cli.ts
 import { Command as Command21 } from "commander";
-import chalk38 from "chalk";
+import chalk39 from "chalk";
 import { readFileSync as readFileSync12 } from "fs";
-import { join as join10, dirname as dirname2 } from "path";
+import { join as join10, dirname as dirname3 } from "path";
 import { fileURLToPath as fileURLToPath2 } from "url";
 
 // src/commands/auth.ts
@@ -1683,12 +1683,61 @@ import { resolve as resolve2 } from "path";
 // src/lib/explorer.ts
 import { readdirSync, readFileSync as readFileSync3, statSync } from "fs";
 import { join as join3, relative, extname, basename } from "path";
-var TEMPLATE_VAR_PATTERN = /\{\{[a-zA-Z_][a-zA-Z0-9_]*\}\}/;
+var DOUBLE_VAR_PATTERN = /\{\{\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*\}\}/g;
+var SINGLE_VAR_PATTERN = /(?<![\\{])\{\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*\}(?!\})/g;
+var FENCED_CODE_PATTERN = /```[\s\S]*?```/g;
 var PROMPT_NAME_PATTERN = /(?:const|let|var|export)\s+\w*(?:prompt|system|agent|instruction|template)\w*\s*=/i;
 var SCHEMA_PATTERN = /["']?(?:inputSchema|outputSchema|properties|required)["']?\s*[=:]/;
 var ZOD_PROMPT_SCHEMA_PATTERN = /(?:const|let|var|export)\s+\w*(?:prompt|output|input|response|schema|message)\w*\s*=\s*z\.object\s*\(/i;
 var PYDANTIC_PATTERN = /class\s+\w+\s*\(\s*(?:BaseModel|BaseSettings)\s*\)/;
 var MARKER_START_PATTERN = /MutagenT:START\s+(\w+)(?:\s+id=(\S+))?/;
+var AGENT_HIGH_PATTERNS = [
+  { pattern: /\bAgentExecutor\b/, reason: "langchain-agent-executor" },
+  { pattern: /\bcreate(?:React|OpenAIFunctions|Tool(?:Calling)?|Structured(?:Chat)?)Agent\b/, reason: "langchain-create-agent" },
+  { pattern: /\bnew\s+Agent\s*\(/, reason: "agent-constructor" },
+  { pattern: /from\s+['"]openai\/agents['"]/, reason: "openai-agents-sdk" },
+  { pattern: /from\s+['"]@openai\/agents['"]/, reason: "openai-agents-sdk" },
+  { pattern: /from\s+['"]crewai['"]/, reason: "crewai-import" },
+  { pattern: /from\s+['"]autogen(?:_agentchat)?['"]/, reason: "autogen-import" },
+  { pattern: /from\s+['"]@?langgraph['"]/, reason: "langgraph-import" },
+  { pattern: /from\s+['"]@langchain\/langgraph['"]/, reason: "langgraph-import" },
+  { pattern: /\bStateGraph\s*\(/, reason: "langgraph-stategraph" }
+];
+var AGENT_MEDIUM_PATTERNS = [
+  { pattern: /\btool_calls\b/, reason: "tool-calls" },
+  { pattern: /\btoolCalls\b/, reason: "tool-calls" },
+  { pattern: /^\s*tools\s*[:=]\s*\[/, reason: "tools-array" },
+  { pattern: /@tool\b/, reason: "tool-decorator" },
+  { pattern: /\bfunction_call\b/, reason: "function-call" },
+  { pattern: /\btool_choice\b/, reason: "tool-choice" }
+];
+function inferPromptVariables(content) {
+  const stripped = content.replace(FENCED_CODE_PATTERN, "");
+  const doubleMatches = [];
+  const singleMatches = [];
+  const dbl = new RegExp(DOUBLE_VAR_PATTERN.source, "g");
+  let m;
+  while ((m = dbl.exec(stripped)) !== null) {
+    if (m[1])
+      doubleMatches.push(m[1]);
+  }
+  const sgl = new RegExp(SINGLE_VAR_PATTERN.source, "g");
+  while ((m = sgl.exec(stripped)) !== null) {
+    if (m[1])
+      singleMatches.push(m[1]);
+  }
+  const delimiter = doubleMatches.length > singleMatches.length ? "double" : "single";
+  const chosen = delimiter === "double" ? doubleMatches : singleMatches;
+  const seen = new Set;
+  const variables = [];
+  for (const v of chosen) {
+    if (!seen.has(v)) {
+      seen.add(v);
+      variables.push(v);
+    }
+  }
+  return { variables, delimiter };
+}
 function walkDir(dir, extensions, excludeDirs, maxDepth, currentDepth = 0) {
   if (currentDepth >= maxDepth)
     return [];
@@ -1728,20 +1777,23 @@ function scanForPrompts(filePath, relativePath) {
   } catch {
     return results;
   }
+  const fileDelimiter = inferPromptVariables(content).delimiter;
   const lines = content.split(`
 `);
   for (let i = 0;i < lines.length; i++) {
     const line = lines[i];
     if (!line)
       continue;
-    if (TEMPLATE_VAR_PATTERN.test(line)) {
+    const inferred = inferPromptVariables(line);
+    if (inferred.variables.length > 0) {
       const preview = line.trim().substring(0, 80);
       results.push({
         file: relativePath,
         line: i + 1,
         preview,
         reason: "template-variable",
-        confidence: "high"
+        confidence: "high",
+        delimiter: fileDelimiter
       });
     }
     if (PROMPT_NAME_PATTERN.test(line)) {
@@ -1752,7 +1804,8 @@ function scanForPrompts(filePath, relativePath) {
           line: i + 1,
           preview,
           reason: "prompt-constant",
-          confidence: "medium"
+          confidence: "medium",
+          delimiter: fileDelimiter
         });
       }
     }
@@ -1764,7 +1817,8 @@ function scanForPrompts(filePath, relativePath) {
           line: i + 1,
           preview,
           reason: "zod-schema",
-          confidence: "low"
+          confidence: "low",
+          delimiter: fileDelimiter
         });
       }
     }
@@ -1776,8 +1830,55 @@ function scanForPrompts(filePath, relativePath) {
           line: i + 1,
           preview,
           reason: "pydantic-model",
+          confidence: "medium",
+          delimiter: fileDelimiter
+        });
+      }
+    }
+  }
+  return results;
+}
+function scanForAgents(filePath, relativePath) {
+  const results = [];
+  let content;
+  try {
+    content = readFileSync3(filePath, "utf-8");
+  } catch {
+    return results;
+  }
+  const lines = content.split(`
+`);
+  const seen = new Set;
+  for (let i = 0;i < lines.length; i++) {
+    const line = lines[i];
+    if (!line)
+      continue;
+    for (const { pattern, reason } of AGENT_HIGH_PATTERNS) {
+      if (pattern.test(line) && !seen.has(i)) {
+        results.push({
+          file: relativePath,
+          line: i + 1,
+          preview: line.trim().substring(0, 80),
+          reason,
+          confidence: "high"
+        });
+        seen.add(i);
+        break;
+      }
+    }
+    if (seen.has(i))
+      continue;
+    for (const { pattern, reason } of AGENT_MEDIUM_PATTERNS) {
+      if (pattern.test(line)) {
+        results.push({
+          file: relativePath,
+          line: i + 1,
+          preview: line.trim().substring(0, 80),
+          reason,
           confidence: "medium"
         });
+        seen.add(i);
+        break;
       }
     }
   }
@@ -1833,7 +1934,8 @@ function scanJsonForSchemas(filePath, relativePath) {
     line: 1,
     preview: preview.substring(0, 80),
     reason: "json-schema",
-    confidence: "medium"
+    confidence: "medium",
+    delimiter: "single"
   });
   return results;
 }
@@ -1890,12 +1992,14 @@ function exploreCodebase(options) {
   const prompts = [];
   const datasets = [];
   const markers = [];
+  const agents = [];
   const sourceFiles = walkDir(rootPath, options.extensions, options.excludeDirs, options.depth);
   for (const filePath of sourceFiles) {
     const relativePath = relative(rootPath, filePath);
     markers.push(...scanForMarkers(filePath, relativePath));
     if (!options.markersOnly) {
       prompts.push(...scanForPrompts(filePath, relativePath));
+      agents.push(...scanForAgents(filePath, relativePath));
     }
   }
   if (!options.markersOnly) {
@@ -1906,7 +2010,7 @@ function exploreCodebase(options) {
     }
     datasets.push(...scanForDatasets(rootPath, rootPath, options.excludeDirs, options.depth));
   }
-  return { prompts, datasets, markers };
+  return { prompts, datasets, markers, agents };
 }
 function parseExtensions(includeGlob) {
   const braceMatch = /\{([^}]+)\}/.exec(includeGlob);
@@ -3539,6 +3643,317 @@ init_sdk_client();
 import { Command as Command3 } from "commander";
 import chalk8 from "chalk";
 init_errors();
+
+// src/commands/prompts/guided-dataset.ts
+init_sdk_client();
+
+// src/commands/prompts/guided-workflow.ts
+init_sdk_client();
+function buildGenericEvalWorkflow() {
+  return {
+    mode: "generic",
+    message: "No prompt-id provided. Follow this sequence to create an evaluation:",
+    workflow: [
+      "1. mutagent explore --json — scan codebase for prompts",
+      "2. mutagent prompts create ... --json — upload the target prompt",
+      "3. Re-run: mutagent prompts evaluation create <prompt-id> --guided --json"
+    ],
+    formatReference: {
+      evalConfig: {
+        criteria: [{
+          name: "<criterion_name>",
+          description: "<rubric: what correct vs incorrect looks like>",
+          evaluationParameter: "<must match a field in inputSchema or outputSchema>"
+        }]
+      }
+    },
+    _directive: {
+      instruction: "Ask the user which prompt they want to evaluate. If unclear, run mutagent explore --json first.",
+      next: ["mutagent explore --json", "mutagent prompts list --json"]
+    }
+  };
+}
+async function checkPromptDisambiguation(promptId, promptName) {
+  try {
+    const client = await getSDKClient();
+    const allPrompts = await client.listPrompts();
+    const targetName = promptName.toLowerCase();
+    const similar = allPrompts.filter((p) => String(p.id) !== promptId && (p.name.toLowerCase().includes(targetName) || targetName.includes(p.name.toLowerCase())));
+    if (similar.length === 0)
+      return null;
+    return {
+      message: `Found ${String(similar.length)} prompt(s) with similar names. Confirm this is the right one:`,
+      selected: { id: promptId, name: promptName },
+      similar: similar.map((p) => ({ id: String(p.id), name: p.name })),
+      askUserQuestions: [{
+        question: `You selected "${promptName}" (ID: ${promptId}). Is this correct, or did you mean one of these?`,
+        options: [
+          `${promptName} (ID: ${promptId}) — SELECTED`,
+          ...similar.map((p) => `${p.name} (ID: ${String(p.id)})`)
+        ]
+      }]
+    };
+  } catch {
+    return null;
+  }
+}
+async function buildGuidedWorkflow(promptId) {
+  const client = await getSDKClient();
+  const prompt = await client.getPrompt(promptId);
+  const truncate = (s, max) => s.length > max ? s.slice(0, max) + "..." : s;
+  const outputProperties = prompt.outputSchema && typeof prompt.outputSchema === "object" ? prompt.outputSchema.properties ?? {} : {};
+  const inputProperties = prompt.inputSchema && typeof prompt.inputSchema === "object" ? prompt.inputSchema.properties ?? {} : {};
+  const inputFields = Object.keys(inputProperties);
+  const outputFields = Object.keys(outputProperties);
+  const allFields = [
+    ...inputFields.map((f) => ({ field: f, source: "inputSchema", fieldSchema: inputProperties[f] })),
+    ...outputFields.map((f) => ({ field: f, source: "outputSchema", fieldSchema: outputProperties[f] }))
+  ];
+  let datasetExample = null;
+  try {
+    const datasets = await client.listDatasets(promptId);
+    if (datasets.length > 0) {
+      try {
+        const dsId = datasets[0]?.id;
+        const items = dsId != null ? await client.listDatasetItems(String(dsId)) : [];
+        if (items.length > 0) {
+          const firstItem = items[0];
+          datasetExample = firstItem?.expectedOutput ?? firstItem?.input ?? null;
+        }
+      } catch {}
+    }
+  } catch {}
+  const askUserQuestions = allFields.map(({ field, source, fieldSchema }) => {
+    const isInput = source === "inputSchema";
+    const question = isInput ? `Define the Minimum Viable Context for "${field}". What data MUST be present in this input for the prompt to produce a correct output? Describe what constitutes complete vs incomplete input, and WHY this field matters.` : `Define what correct "${field}" looks like. What structure, content, or qualities make it good vs bad? Give concrete examples of good and bad outputs.`;
+    const hint = isInput ? `Input fields define what data the prompt NEEDS to work correctly. Without defining minimum viable context, the optimizer cannot detect whether failures come from bad input or bad prompt logic.` : null;
+    const header = isInput ? `${field} [INPUT]` : `${field} [OUTPUT]`;
+    const defineDesc = isInput ? `Describe what data MUST be present in "${field}" and WHY the prompt needs it. Focus on completeness and minimum viable context.` : `Describe what a correct vs incorrect "${field}" looks like. Focus on structure, content quality, and concrete examples — not numeric scores.`;
+    return {
+      question,
+      header,
+      ...hint != null ? { hint } : {},
+      options: [
+        {
+          label: "Define rubric",
+          description: defineDesc
+        },
+        {
+          label: "See suggestion",
+          description: `Get a suggested rubric based on the prompt and schema definition for "${field}". You can refine it.`
+        }
+      ],
+      multiSelect: false,
+      context: {
+        fieldType: fieldSchema?.type ?? "unknown",
+        fieldDescription: fieldSchema?.description ?? null,
+        fieldSource: source,
+        promptExcerpt: truncate(prompt.humanPrompt ?? prompt.systemPrompt ?? prompt.rawPrompt ?? "", 200),
+        exampleValue: datasetExample?.[field] ?? null
+      }
+    };
+  });
+  const result = {
+    prompt: { id: promptId, name: prompt.name },
+    inputSchema: { fields: inputFields },
+    outputSchema: { fields: outputFields },
+    workflow: {
+      description: "Follow these steps to create an evaluation for this prompt:",
+      steps: [
+        { step: 1, action: "Review prompt schemas and context", detail: "Understand what the prompt does and what each field means." },
+        { step: 2, action: "Define correctness criteria for EVERY field", detail: `For EACH of the ${String(allFields.length)} fields in askUserQuestions, use AskUserQuestion to collect a rubric. INPUT fields (${String(inputFields.length)}): collect minimum viable context definitions — what data MUST exist. OUTPUT fields (${String(outputFields.length)}): collect correctness criteria — what correct vs incorrect looks like. Do NOT skip any field. Use the context hints to suggest rubrics when asked.` },
+        { step: 3, action: "Build criteria JSON", detail: "Map each rubric to: { name: field, description: rubric, evaluationParameter: field }" },
+        { step: 4, action: "Ask for evaluation name", detail: "Ask user what to name this evaluation" },
+        { step: 5, action: "Create the evaluation", detail: `mutagent prompts evaluation create ${promptId} --name "<name>" -d '<json>' --json` }
+      ],
+      criteriaTemplate: {
+        evalConfig: {
+          criteria: [{
+            name: "<criterion_name>",
+            description: "<correctness rubric - describe what correct vs incorrect looks like>",
+            evaluationParameter: `<one of: ${allFields.length > 0 ? allFields.map((f) => f.field).join(", ") : "no fields detected - check prompt schemas"}>`
+          }]
+        }
+      },
+      exampleCommand: `mutagent prompts evaluation create ${promptId} --name "My Evaluation" -d '${JSON.stringify({
+        evalConfig: {
+          criteria: outputFields.slice(0, 1).map((f) => ({
+            name: `${f} accuracy`,
+            description: `Correct if ${f} matches expected structure and content; incorrect if missing, malformed, or off-topic`,
+            evaluationParameter: f
+          }))
+        }
+      })}'`
+    },
+    askUserQuestions,
+    _directive: {
+      display: "workflow_guide",
+      instruction: `MANDATORY ORDERS — READ CAREFULLY:
+` + `1. You MUST ask the user about EVERY SINGLE field listed in askUserQuestions. There are ${String(allFields.length)} fields total.
+` + `2. Do NOT skip ANY field. Do NOT decide that a field "doesn't need" a definition.
+` + `3. INPUT fields (${String(inputFields.length)}): These define MINIMUM VIABLE CONTEXT — the data that MUST exist in the input for the prompt to work. Ask: "What data MUST be present in this field and WHY does the prompt need it?"
+` + `   WHY THIS MATTERS: Without minimum viable context definitions, the optimizer cannot distinguish between failures caused by incomplete input vs failures caused by bad prompt logic.
+` + `4. OUTPUT fields (${String(outputFields.length)}): These define correctness criteria. Ask: "What does correct vs incorrect for this field look like?"
+` + `5. After collecting ALL ${String(allFields.length)} definitions, construct the --data JSON with criteria for EVERY field.
+` + `6. VIOLATION: Skipping ANY field or telling the user a field doesn't need a definition is a protocol violation.
+` + `
+` + `7. EVAL CRITERIA FRAMING — INPUT vs OUTPUT scoping (mirror of SKILL.md → "Guided Eval Criteria"):
+` + `   INPUT fields → Minimum Viable Context (MVC).
+` + `     Ask: "Is the data required for the prompt to succeed actually present in this {variable}?"
+` + `     Rubric uses a COMPLETENESS scale grounded in observable presence of required data:
+` + `       1.0 = all required context present, no ambiguity
+` + `       0.5 = some context present, enough to attempt but likely partial/hedged answer
+` + `       0.0 = critical context missing, prompt cannot succeed regardless of model
+` + `   OUTPUT fields → Output Standards.
+` + `     Ask: "Does the response meet content correctness + structural correctness?"
+` + `     Rubric uses a CORRECTNESS scale with concrete pass/fail tiers, e.g.:
+` + `       1.0 = valid shape, all required fields, faithful content
+` + `       0.5 = valid shape, 1-2 fields hedged or partially correct
+` + `       0.0 = invalid shape OR fabricated facts
+` + `   NEVER use vague rubrics like "0-1 scale" or "score based on accuracy".
+` + `   Canonical source: .claude/skills/mutagent-cli/SKILL.md → "Guided Eval Criteria".
+` + "   Note: template variables use single-brace `{variable}` (platform canonical). `mutagent explore`\n" + "   infers single vs double per-file via inferPromptVariables().",
+      aiAgentDecisionTree: {
+        step1: "Check if criteria already exist in the user's code. If criteria match expected shape (name + description + evaluationParameter targeting schema fields), use --data directly.",
+        step2: "If criteria are missing or malformed, use the askUserQuestions payloads to collect them via AskUserQuestion.",
+        step3: `After collecting all criteria, construct the --data JSON and run: mutagent prompts evaluation create <prompt-id> --name "<name>" -d '<json>'`
+      }
+    },
+    _meta: { inputFields, outputFields, allFields }
+  };
+  const disambiguation = await checkPromptDisambiguation(promptId, prompt.name);
+  if (disambiguation) {
+    result.disambiguation = disambiguation;
+  }
+  return result;
+}
+
+// src/commands/prompts/guided-dataset.ts
+function buildGenericDatasetWorkflow() {
+  return {
+    mode: "generic",
+    message: "No prompt-id provided. Follow this sequence to create a dataset:",
+    workflow: [
+      "1. mutagent explore --json — scan codebase for prompts",
+      "2. mutagent prompts create ... --json — upload the target prompt",
+      "3. Re-run: mutagent prompts dataset add <prompt-id> --guided --json"
+    ],
+    formatReference: {
+      item: {
+        input: { "<inputSchema_field>": "<value>" },
+        expectedOutput: { "<outputSchema_field>": "<expected_value>" }
+      }
+    },
+    suggestedCategories: [
+      { name: "Edge Cases", description: "Boundary inputs, empty fields, malformed data", priority: "high" },
+      { name: "Hard Cases", description: "Ambiguous inputs, adversarial examples", priority: "high" },
+      { name: "Representative Cases", description: "Typical production inputs", priority: "medium" }
+    ],
+    _directive: {
+      instruction: "Ask the user which prompt they want to create a dataset for. If unclear, run mutagent explore --json first.",
+      next: ["mutagent explore --json", "mutagent prompts list --json"]
+    }
+  };
+}
+function buildGuidedDatasetWorkflow(prompt) {
+  const promptId = String(prompt.id);
+  const outputProperties = prompt.outputSchema && typeof prompt.outputSchema === "object" ? prompt.outputSchema.properties ?? {} : {};
+  const inputProperties = prompt.inputSchema && typeof prompt.inputSchema === "object" ? prompt.inputSchema.properties ?? {} : {};
+  const inputFields = Object.keys(inputProperties);
+  const outputFields = Object.keys(outputProperties);
+  const templateInput = {};
+  for (const f of inputFields) {
+    templateInput[f] = "<value>";
+  }
+  const templateOutput = {};
+  for (const f of outputFields) {
+    templateOutput[f] = "<expected>";
+  }
+  const askUserQuestions = [
+    {
+      field: "_general",
+      source: "inputSchema",
+      question: "What are the hardest inputs for this prompt? Describe scenarios that typically cause failures or ambiguous outputs."
+    },
+    {
+      field: "_edge_cases",
+      source: "inputSchema",
+      question: "What edge cases have caused failures in production? Include boundary inputs, empty fields, and malformed data."
+    },
+    ...inputFields.map((f) => ({
+      field: f,
+      source: "inputSchema",
+      question: `What values should "${f}" have in test cases? Include edge cases (empty, unicode, very long, malformed).`
+    })),
+    ...outputFields.map((f) => ({
+      field: f,
+      source: "outputSchema",
+      question: `What should correct "${f}" look like? Describe the format, tone, completeness criteria, and examples of incorrect outputs.`
+    }))
+  ];
+  const suggestedCategories = [
+    {
+      name: "Edge Cases",
+      description: "Boundary inputs, empty fields, malformed data",
+      priority: "high"
+    },
+    {
+      name: "Hard Cases",
+      description: "Ambiguous inputs, adversarial examples, domain traps",
+      priority: "high"
+    },
+    {
+      name: "Representative Cases",
+      description: "Typical production inputs",
+      priority: "medium"
+    }
+  ];
+  const totalSchemaFields = inputFields.length + outputFields.length;
+  const hasSchema = totalSchemaFields > 0;
+  return {
+    promptId,
+    promptName: prompt.name,
+    schemaFields: {
+      input: inputFields,
+      output: outputFields
+    },
+    suggestedCategories,
+    templateItem: {
+      input: hasSchema ? templateInput : { "<field>": "<value>" },
+      expectedOutput: hasSchema ? templateOutput : { "<field>": "<expected>" }
+    },
+    askUserQuestions,
+    guidance: {
+      minItems: 5,
+      priorityRule: "Hard cases that expose prompt weaknesses > easy cases that always pass",
+      steps: [
+        "Collect answers to each question from the user",
+        "Generate 5-10 dataset items covering all categories",
+        "Ensure at least 2 hard/edge cases per category",
+        'Format as JSON array: [{"input": {...}, "expectedOutput": {...}}, ...]',
+        `Run: mutagent prompts dataset add ${promptId} -d '<json>' --name '<name>' --json`
+      ]
+    },
+    _directive: {
+      instruction: "Collect domain-specific dataset items from the user. " + "Ask about each schema field. " + "Prioritize hard cases over easy cases. " + "Then construct the dataset JSON and upload.",
+      next: [
+        `mutagent prompts dataset add ${promptId} -d '<constructed-json>' --name '<name>' --json`
+      ]
+    }
+  };
+}
+async function fetchAndBuildGuidedDatasetWorkflow(promptId) {
+  const client = await getSDKClient();
+  const prompt = await client.getPrompt(promptId);
+  const result = buildGuidedDatasetWorkflow(prompt);
+  const disambiguation = await checkPromptDisambiguation(promptId, prompt.name);
+  if (disambiguation) {
+    result.disambiguation = disambiguation;
+  }
+  return result;
+}
+
+// src/commands/prompts/datasets.ts
 function registerDatasetCommands(prompts) {
   const dataset = new Command3("dataset").description("Manage datasets for prompts").addHelpText("after", `
 Examples:
@@ -3638,11 +4053,19 @@ Verify the dataset ID exists, or list datasets for a prompt to find valid IDs.`)
       }
     }
   });
-  dataset.command("add").description("Add dataset to a prompt").argument("<prompt-id>", "Prompt ID (from: mutagent prompts list)").option("-d, --data <json>", "Inline JSON array of dataset items").option("-n, --name <name>", "Dataset name").addHelpText("after", `
+  dataset.command("add").description("Add dataset to a prompt").argument("[prompt-id]", "Prompt ID (from: mutagent prompts list) — optional when --guided is set").option("-d, --data <json>", "Inline JSON array of dataset items").option("-n, --name <name>", "Dataset name").option("--guided", "Guided mode — analyze prompt schema and suggest dataset categories. prompt-id is optional when --guided is set.").addHelpText("after", `
 Examples:
+  ${chalk8.dim("$")} mutagent prompts dataset add --guided --json                 # no prompt-id: returns generic workflow + upload instructions
+  ${chalk8.dim("$")} mutagent prompts dataset add <prompt-id> --guided --json    # recommended: schema-aware category suggestions
   ${chalk8.dim("$")} mutagent prompts dataset add <prompt-id> -d '[{"input":{"text":"hello"},"expectedOutput":{"result":"world"}}]'
   ${chalk8.dim("$")} mutagent prompts dataset add <prompt-id> -d '[{"input":{"text":"hello"},"expectedOutput":{"result":"world"}}]' --name "My Dataset"
 
+Guided mode (--guided):
+  Fetches the prompt's inputSchema + outputSchema and returns structured JSON
+  that guides a coding agent to collect domain-specific dataset items.
+  The agent uses the guidance to ask the user questions, then constructs
+  the dataset JSON and calls: mutagent prompts dataset add <id> -d '<json>' --name '<name>' --json
+
 Inline data format (-d):
   JSON array of objects, e.g.:
   ${chalk8.dim('[{"input": {"text": "hello"}, "expectedOutput": {"result": "world"}}]')}
@@ -3657,11 +4080,26 @@ ${chalk8.yellow("AI Agent (MANDATORY):")}
   expectedOutput is REQUIRED for evaluation scoring.
   Check schemas: mutagent prompts get <prompt-id> --json
 
-${chalk8.red("Required: --data must be provided.")}
+${chalk8.red("Required: --data or --guided must be provided.")}
     `).action(async (promptId, options) => {
-    const isJson = getJsonFlag(prompts);
+    const isJson = options.guided ? true : getJsonFlag(prompts);
     const output = new OutputFormatter(isJson ? "json" : "table");
     try {
+      if (options.guided && !promptId) {
+        const result = buildGenericDatasetWorkflow();
+        output.output(result);
+        return;
+      }
+      if (options.guided) {
+        const guidedResult = await fetchAndBuildGuidedDatasetWorkflow(promptId);
+        output.output(guidedResult);
+        return;
+      }
+      if (!promptId) {
+        throw new MutagentError("MISSING_ARGUMENTS", "prompt-id is required", `Run: mutagent prompts dataset add --help
+` + `Usage: mutagent prompts dataset add <prompt-id> [options]
+` + "With --guided: mutagent prompts dataset add --guided --json (returns generic workflow)");
+      }
       if (!options.data) {
         throw new MutagentError("MISSING_ARGUMENTS", "-d/--data is required", `Run: mutagent prompts dataset add --help
 ` + "Use -d '[{...}]' to provide inline JSON data");
@@ -3704,20 +4142,21 @@ Each item must have: {"input": {...}, "expectedOutput": {...}}`);
       if (!datasetName) {
         throw new MutagentError("MISSING_ARGUMENTS", "Dataset name is required", `Run: mutagent prompts dataset add --help
 ` + `[Agent: Ask the user for a dataset name via AskUserQuestion, then pass --name]
-` + `Use --name <name>, e.g., mutagent prompts dataset add ${promptId} --name "my-dataset" -d '[...]'`);
+` + `Use --name <name>, e.g., mutagent prompts dataset add ${promptId ?? "<prompt-id>"} --name "my-dataset" -d '[...]'`);
       }
+      const resolvedPromptId = promptId;
       const client = await getSDKClient();
-      const datasetResult = await client.addDataset(promptId, content, datasetName);
+      const datasetResult = await client.addDataset(resolvedPromptId, content, datasetName);
       if (isJson) {
         let rsState;
         try {
-          const evals = await client.listEvaluations(promptId);
+          const evals = await client.listEvaluations(resolvedPromptId);
           rsState = { evaluations: evals.length };
         } catch {}
-        const directive = datasetAddedDirective(promptId, datasetResult.id, datasetResult.name, datasetResult.itemCount, rsState);
+        const directive = datasetAddedDirective(resolvedPromptId, datasetResult.id, datasetResult.name, datasetResult.itemCount, rsState);
         output.output({
           ...datasetResult,
-          _links: datasetLinks(promptId, datasetResult.id),
+          _links: datasetLinks(resolvedPromptId, datasetResult.id),
           _directive: directive
         });
         echoDirectiveToStderr(directive);
@@ -3725,7 +4164,7 @@ Each item must have: {"input": {...}, "expectedOutput": {...}}`);
         for (const w of warnings) {
           output.warn(w);
         }
-        output.success(`Added dataset "${datasetResult.name}" to prompt: ${promptId} (id: ${String(datasetResult.id)})`);
+        output.success(`Added dataset "${datasetResult.name}" to prompt: ${resolvedPromptId} (id: ${String(datasetResult.id)})`);
         if (datasetResult.itemCount !== undefined && datasetResult.itemCount > 0) {
           output.info(`Items uploaded: ${String(datasetResult.itemCount)}`);
         }
@@ -3736,14 +4175,14 @@ Each item must have: {"input": {...}, "expectedOutput": {...}}`);
           resourceType: "Dataset",
           id: datasetResult.id,
           name: datasetResult.name,
-          dashboardUrl: datasetLink(promptId, datasetResult.id),
-          apiPath: `/api/prompts/${promptId}/datasets/${String(datasetResult.id)}`
+          dashboardUrl: datasetLink(resolvedPromptId, datasetResult.id),
+          apiPath: `/api/prompts/${resolvedPromptId}/datasets/${String(datasetResult.id)}`
         });
         console.log(hints);
       }
       updateMutationContext((ctx) => {
         ctx.addDiscoveredDataset("inline-data", datasetResult.name, datasetResult.itemCount ?? 0);
-        ctx.markDatasetUploaded("inline-data", String(datasetResult.id), promptId);
+        ctx.markDatasetUploaded("inline-data", String(datasetResult.id), resolvedPromptId);
       });
     } catch (error) {
       handleError(error, isJson);
@@ -3883,144 +4322,33 @@ async function resolveNumericPromptId(client, promptGroupId) {
   return sorted[0]?.id ?? null;
 }
 
-// src/commands/prompts/guided-workflow.ts
-init_sdk_client();
-async function buildGuidedWorkflow(promptId) {
-  const client = await getSDKClient();
-  const prompt = await client.getPrompt(promptId);
-  const truncate = (s, max) => s.length > max ? s.slice(0, max) + "..." : s;
-  const outputProperties = prompt.outputSchema && typeof prompt.outputSchema === "object" ? prompt.outputSchema.properties ?? {} : {};
-  const inputProperties = prompt.inputSchema && typeof prompt.inputSchema === "object" ? prompt.inputSchema.properties ?? {} : {};
-  const inputFields = Object.keys(inputProperties);
-  const outputFields = Object.keys(outputProperties);
-  const allFields = [
-    ...inputFields.map((f) => ({ field: f, source: "inputSchema", fieldSchema: inputProperties[f] })),
-    ...outputFields.map((f) => ({ field: f, source: "outputSchema", fieldSchema: outputProperties[f] }))
-  ];
-  let datasetExample = null;
-  try {
-    const datasets = await client.listDatasets(promptId);
-    if (datasets.length > 0) {
-      try {
-        const dsId = datasets[0]?.id;
-        const items = dsId != null ? await client.listDatasetItems(String(dsId)) : [];
-        if (items.length > 0) {
-          const firstItem = items[0];
-          datasetExample = firstItem?.expectedOutput ?? firstItem?.input ?? null;
-        }
-      } catch {}
-    }
-  } catch {}
-  const askUserQuestions = allFields.map(({ field, source, fieldSchema }) => {
-    const isInput = source === "inputSchema";
-    const question = isInput ? `Define the Minimum Viable Context for "${field}". What data MUST be present in this input for the prompt to produce a correct output? Describe what constitutes complete vs incomplete input, and WHY this field matters.` : `Define what correct "${field}" looks like. What structure, content, or qualities make it good vs bad? Give concrete examples of good and bad outputs.`;
-    const hint = isInput ? `Input fields define what data the prompt NEEDS to work correctly. Without defining minimum viable context, the optimizer cannot detect whether failures come from bad input or bad prompt logic.` : null;
-    const header = isInput ? `${field} [INPUT]` : `${field} [OUTPUT]`;
-    const defineDesc = isInput ? `Describe what data MUST be present in "${field}" and WHY the prompt needs it. Focus on completeness and minimum viable context.` : `Describe what a correct vs incorrect "${field}" looks like. Focus on structure, content quality, and concrete examples — not numeric scores.`;
-    return {
-      question,
-      header,
-      ...hint != null ? { hint } : {},
-      options: [
-        {
-          label: "Define rubric",
-          description: defineDesc
-        },
-        {
-          label: "See suggestion",
-          description: `Get a suggested rubric based on the prompt and schema definition for "${field}". You can refine it.`
-        }
-      ],
-      multiSelect: false,
-      context: {
-        fieldType: fieldSchema?.type ?? "unknown",
-        fieldDescription: fieldSchema?.description ?? null,
-        fieldSource: source,
-        promptExcerpt: truncate(prompt.humanPrompt ?? prompt.systemPrompt ?? prompt.rawPrompt ?? "", 200),
-        exampleValue: datasetExample?.[field] ?? null
-      }
-    };
-  });
-  return {
-    prompt: { id: promptId, name: prompt.name },
-    inputSchema: { fields: inputFields },
-    outputSchema: { fields: outputFields },
-    workflow: {
-      description: "Follow these steps to create an evaluation for this prompt:",
-      steps: [
-        { step: 1, action: "Review prompt schemas and context", detail: "Understand what the prompt does and what each field means." },
-        { step: 2, action: "Define correctness criteria for EVERY field", detail: `For EACH of the ${String(allFields.length)} fields in askUserQuestions, use AskUserQuestion to collect a rubric. INPUT fields (${String(inputFields.length)}): collect minimum viable context definitions — what data MUST exist. OUTPUT fields (${String(outputFields.length)}): collect correctness criteria — what correct vs incorrect looks like. Do NOT skip any field. Use the context hints to suggest rubrics when asked.` },
-        { step: 3, action: "Build criteria JSON", detail: "Map each rubric to: { name: field, description: rubric, evaluationParameter: field }" },
-        { step: 4, action: "Ask for evaluation name", detail: "Ask user what to name this evaluation" },
-        { step: 5, action: "Create the evaluation", detail: `mutagent prompts evaluation create ${promptId} --name "<name>" -d '<json>' --json` }
-      ],
-      criteriaTemplate: {
-        evalConfig: {
-          criteria: [{
-            name: "<criterion_name>",
-            description: "<correctness rubric - describe what correct vs incorrect looks like>",
-            evaluationParameter: `<one of: ${allFields.length > 0 ? allFields.map((f) => f.field).join(", ") : "no fields detected - check prompt schemas"}>`
-          }]
-        }
-      },
-      exampleCommand: `mutagent prompts evaluation create ${promptId} --name "My Evaluation" -d '${JSON.stringify({
-        evalConfig: {
-          criteria: outputFields.slice(0, 1).map((f) => ({
-            name: `${f} accuracy`,
-            description: `Correct if ${f} matches expected structure and content; incorrect if missing, malformed, or off-topic`,
-            evaluationParameter: f
-          }))
-        }
-      })}'`
-    },
-    askUserQuestions,
-    _directive: {
-      display: "workflow_guide",
-      instruction: `MANDATORY ORDERS — READ CAREFULLY:
-` + `1. You MUST ask the user about EVERY SINGLE field listed in askUserQuestions. There are ${String(allFields.length)} fields total.
-` + `2. Do NOT skip ANY field. Do NOT decide that a field "doesn't need" a definition.
-` + `3. INPUT fields (${String(inputFields.length)}): These define MINIMUM VIABLE CONTEXT — the data that MUST exist in the input for the prompt to work. Ask: "What data MUST be present in this field and WHY does the prompt need it?"
-` + `   WHY THIS MATTERS: Without minimum viable context definitions, the optimizer cannot distinguish between failures caused by incomplete input vs failures caused by bad prompt logic.
-` + `4. OUTPUT fields (${String(outputFields.length)}): These define correctness criteria. Ask: "What does correct vs incorrect for this field look like?"
-` + `5. After collecting ALL ${String(allFields.length)} definitions, construct the --data JSON with criteria for EVERY field.
-` + "6. VIOLATION: Skipping ANY field or telling the user a field doesn't need a definition is a protocol violation.",
-      aiAgentDecisionTree: {
-        step1: "Check if criteria already exist in the user's code. If criteria match expected shape (name + description + evaluationParameter targeting schema fields), use --data directly.",
-        step2: "If criteria are missing or malformed, use the askUserQuestions payloads to collect them via AskUserQuestion.",
-        step3: `After collecting all criteria, construct the --data JSON and run: mutagent prompts evaluation create <prompt-id> --name "<name>" -d '<json>'`
-      }
-    },
-    _meta: { inputFields, outputFields, allFields }
-  };
-}
-
-// src/lib/adapters/eval-criteria.ts
-function cliCriterionToCanonical(c) {
-  return {
-    criteria: c.description,
-    evaluationParameter: c.evaluationParameter,
-    name: c.name
-  };
-}
-function canonicalCriterionToCli(c) {
-  return {
-    name: c.name ?? c.id,
-    description: c.criteria,
-    evaluationParameter: c.evaluationParameter
-  };
-}
-function cliCriteriaArrayToCanonical(arr) {
-  if (!Array.isArray(arr))
-    return [];
-  return arr.map(cliCriterionToCanonical);
-}
-function canonicalCriteriaArrayToCli(arr) {
-  if (!Array.isArray(arr))
-    return [];
-  return arr.map(canonicalCriterionToCli);
-}
-
-// src/commands/prompts/evaluations-run.ts
+// src/lib/adapters/eval-criteria.ts
+function cliCriterionToCanonical(c) {
+  return {
+    criteria: c.description,
+    evaluationParameter: c.evaluationParameter,
+    name: c.name
+  };
+}
+function canonicalCriterionToCli(c) {
+  return {
+    name: c.name ?? c.id,
+    description: c.criteria,
+    evaluationParameter: c.evaluationParameter
+  };
+}
+function cliCriteriaArrayToCanonical(arr) {
+  if (!Array.isArray(arr))
+    return [];
+  return arr.map(cliCriterionToCanonical);
+}
+function canonicalCriteriaArrayToCli(arr) {
+  if (!Array.isArray(arr))
+    return [];
+  return arr.map(canonicalCriterionToCli);
+}
+
+// src/commands/prompts/evaluations-run.ts
 init_sdk_client();
 import chalk9 from "chalk";
 init_errors();
@@ -4213,8 +4541,9 @@ Examples:
       handleError(error, isJson);
     }
   });
-  evaluation.command("create").description("Create an evaluation configuration for a prompt").argument("<prompt-id>", "Prompt ID (from: mutagent prompts list)").option("-d, --data <json>", "Evaluation as JSON string (for pre-validated criteria only)").option("-n, --name <name>", "Evaluation name (required unless --guided)").option("--description <text>", "Evaluation description").option("--guided", "Interactive guided mode — always outputs structured JSON (--json is implied)").addHelpText("after", `
+  evaluation.command("create").description("Create an evaluation configuration for a prompt").argument("[prompt-id]", "Prompt ID (from: mutagent prompts list) — optional when --guided is set").option("-d, --data <json>", "Evaluation as JSON string (for pre-validated criteria only)").option("-n, --name <name>", "Evaluation name (required unless --guided)").option("--description <text>", "Evaluation description").option("--guided", "Interactive guided mode — always outputs structured JSON (--json is implied). prompt-id is optional when --guided is set.").addHelpText("after", `
 Examples:
+  ${chalk10.dim("$")} mutagent prompts evaluation create --guided --json               ${chalk10.dim("# no prompt-id: returns generic workflow + upload instructions")}
   ${chalk10.dim("$")} mutagent prompts evaluation create <prompt-id> --guided          ${chalk10.dim("# recommended: shows workflow guide + schema fields")}
   ${chalk10.dim("$")} mutagent prompts evaluation create <prompt-id> --guided --json  ${chalk10.dim("# structured workflow for AI agents")}
   ${chalk10.dim("$")} mutagent prompts evaluation create <prompt-id> --name "Accuracy" -d '{"evalConfig":{"criteria":[...]}}'  ${chalk10.dim("# power user")}
@@ -4251,6 +4580,16 @@ ${chalk10.dim("Get prompt IDs: mutagent prompts list")}
     }
     const output = new OutputFormatter(isJson ? "json" : "table");
     try {
+      if (options.guided && !promptId) {
+        const result = buildGenericEvalWorkflow();
+        output.output(result);
+        return;
+      }
+      if (!promptId) {
+        throw new MutagentError("MISSING_ARGUMENTS", "prompt-id is required", `Run: mutagent prompts evaluation create --help
+` + `Usage: mutagent prompts evaluation create <prompt-id> [options]
+` + "With --guided: mutagent prompts evaluation create --guided --json (returns generic workflow)");
+      }
       const renderGuidedWorkflow = (wf) => {
         const { inputFields, outputFields, allFields } = wf._meta;
         if (isJson) {
@@ -7238,18 +7577,32 @@ AI AGENT: Execute these steps using Bash for install, Write/Edit for code change
 
 // src/commands/agents/index.ts
 import { Command as Command10 } from "commander";
-import chalk25 from "chalk";
+import chalk26 from "chalk";
 
 // src/commands/agents/agents-crud.ts
 init_sdk_client();
-import chalk24 from "chalk";
+import chalk25 from "chalk";
 init_errors();
+
+// src/lib/agents-wip-banner.ts
+import chalk24 from "chalk";
+var AGENTS_WIP_BANNER = `${chalk24.yellow("Agents: Early Access / WIP")}
+  ${chalk24.bold("Agent optimization & evaluation are actively in development.")}
+  ${chalk24.dim("The Prompt path (single-shot LLM calls with output schemas) is production-ready.")}
+  ${chalk24.dim("For Agents (multi-turn / tool-calling), join early access:")}
+  ${chalk24.cyan("https://www.mutagent.io/agents-partnership")}
+  ${chalk24.dim("`mutagent agents ...` CRUD commands exist but optimization/eval are NOT ready.")}`;
+var AGENTS_WIP_BANNER_SHORT = `${chalk24.yellow("[Agents WIP]")} ${chalk24.dim("CRUD only — optimization & evaluation in development.")} ${chalk24.cyan("https://www.mutagent.io/agents-partnership")}`;
+
+// src/commands/agents/agents-crud.ts
 function registerAgentsCrud(agents) {
   agents.command("list").description("List all agents").option("-l, --limit <n>", "Limit results", "50").option("-o, --offset <n>", "Offset for pagination").option("-n, --name <name>", "Filter by name").option("-s, --status <status>", "Filter by status (active, paused, archived)").addHelpText("after", `
+${AGENTS_WIP_BANNER_SHORT}
+
 Examples:
-  ${chalk24.dim("$")} mutagent agents list
-  ${chalk24.dim("$")} mutagent agents list --status active
-  ${chalk24.dim("$")} mutagent agents list --name "reviewer" --json
+  ${chalk25.dim("$")} mutagent agents list
+  ${chalk25.dim("$")} mutagent agents list --status active
+  ${chalk25.dim("$")} mutagent agents list --name "reviewer" --json
     `).action(async (options) => {
     const isJson = getJsonFlag(agents);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -7298,9 +7651,11 @@ Examples:
     }
   });
   agents.command("get").description("Get agent details").argument("<id>", "Agent ID").addHelpText("after", `
+${AGENTS_WIP_BANNER_SHORT}
+
 Examples:
-  ${chalk24.dim("$")} mutagent agents get <agent-id>
-  ${chalk24.dim("$")} mutagent agents get <agent-id> --json
+  ${chalk25.dim("$")} mutagent agents get <agent-id>
+  ${chalk25.dim("$")} mutagent agents get <agent-id> --json
     `).action(async (id) => {
     const isJson = getJsonFlag(agents);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -7330,11 +7685,11 @@ Examples:
         };
         output.output(formatted);
         if (agent.systemPrompt) {
-          console.log(chalk24.bold(`
+          console.log(chalk25.bold(`
 System Prompt:`));
-          console.log(chalk24.gray("─".repeat(60)));
+          console.log(chalk25.gray("─".repeat(60)));
           console.log(agent.systemPrompt);
-          console.log(chalk24.gray("─".repeat(60)));
+          console.log(chalk25.gray("─".repeat(60)));
         }
       }
     } catch (error) {
@@ -7342,18 +7697,20 @@ System Prompt:`));
     }
   });
   agents.command("create").description("Create a new agent").option("-d, --data <json>", "Agent as JSON string (recommended for CI/scripts/agents)").option("-n, --name <name>", "Agent name").option("-s, --slug <slug>", "Agent slug (URL-friendly identifier)").option("-p, --system-prompt <prompt>", "System prompt").option("-m, --model <model>", "Model (claude-sonnet-4-5, claude-opus-4-5, claude-haiku-4-5)").option("--description <desc>", "Agent description").addHelpText("after", `
+${AGENTS_WIP_BANNER_SHORT}
+
 Examples:
-  ${chalk24.dim("$")} mutagent agents create --name "Code Reviewer" --slug code-reviewer --system-prompt "You are a code reviewer..."
-  ${chalk24.dim("$")} mutagent agents create -d '{"name":"Code Reviewer","slug":"code-reviewer","systemPrompt":"You are a code reviewer..."}'
+  ${chalk25.dim("$")} mutagent agents create --name "Code Reviewer" --slug code-reviewer --system-prompt "You are a code reviewer..."
+  ${chalk25.dim("$")} mutagent agents create -d '{"name":"Code Reviewer","slug":"code-reviewer","systemPrompt":"You are a code reviewer..."}'
 
 Expected JSON (--data):
-  ${chalk24.dim('{"name":"<name>","slug":"<slug>","systemPrompt":"<system prompt>","model":"<model-id>","description":"<description>"}')}
+  ${chalk25.dim('{"name":"<name>","slug":"<slug>","systemPrompt":"<system prompt>","model":"<model-id>","description":"<description>"}')}
 
 Input Methods (pick one, priority order):
-  --name/--slug/...  Individual flags ${chalk24.green("(recommended)")}
+  --name/--slug/...  Individual flags ${chalk25.green("(recommended)")}
   -d, --data         Inline JSON object (CI/scripts/agents)
 
-${chalk24.red("Required: name, slug, systemPrompt.")} ${chalk24.dim("CLI flags override --data fields.")}
+${chalk25.red("Required: name, slug, systemPrompt.")} ${chalk25.dim("CLI flags override --data fields.")}
     `).action(async (options) => {
     const isJson = getJsonFlag(agents);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -7398,16 +7755,18 @@ ${chalk24.red("Required: name, slug, systemPrompt.")} ${chalk24.dim("CLI flags o
     }
   });
   agents.command("update").description("Update an agent").argument("<id>", "Agent ID").option("-d, --data <json>", "Agent updates as JSON string (CI/scripts/agents)").option("-n, --name <name>", "New name").option("-p, --system-prompt <prompt>", "New system prompt").option("-m, --model <model>", "New model").option("--description <desc>", "New description").option("-s, --status <status>", "New status (active, paused, archived)").addHelpText("after", `
+${AGENTS_WIP_BANNER_SHORT}
+
 Examples:
-  ${chalk24.dim("$")} mutagent agents update <id> --name "New Name"
-  ${chalk24.dim("$")} mutagent agents update <id> --system-prompt "Updated prompt" --status active
-  ${chalk24.dim("$")} mutagent agents update <id> -d '{"name":"New Name","systemPrompt":"Updated prompt"}'
+  ${chalk25.dim("$")} mutagent agents update <id> --name "New Name"
+  ${chalk25.dim("$")} mutagent agents update <id> --system-prompt "Updated prompt" --status active
+  ${chalk25.dim("$")} mutagent agents update <id> -d '{"name":"New Name","systemPrompt":"Updated prompt"}'
 
 Input Methods (pick one, priority order):
-  --name/--system-prompt/...  Individual flags ${chalk24.green("(recommended)")}
+  --name/--system-prompt/...  Individual flags ${chalk25.green("(recommended)")}
   -d, --data                  Inline JSON object (CI/scripts/agents)
 
-${chalk24.dim("CLI flags override --data fields.")}
+${chalk25.dim("CLI flags override --data fields.")}
     `).action(async (id, options) => {
     const isJson = getJsonFlag(agents);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -7454,12 +7813,14 @@ ${chalk24.dim("CLI flags override --data fields.")}
     }
   });
   agents.command("delete").description("Delete an agent").argument("<id>", "Agent ID").option("--force", "Skip confirmation").addHelpText("after", `
+${AGENTS_WIP_BANNER_SHORT}
+
 Examples:
-  ${chalk24.dim("$")} mutagent agents delete <id>
-  ${chalk24.dim("$")} mutagent agents delete <id> --force
-  ${chalk24.dim("$")} mutagent agents delete <id> --force --json
+  ${chalk25.dim("$")} mutagent agents delete <id>
+  ${chalk25.dim("$")} mutagent agents delete <id> --force
+  ${chalk25.dim("$")} mutagent agents delete <id> --force --json
 
-${chalk24.dim("Tip: Use --force to skip confirmation (required for non-interactive/CI usage).")}
+${chalk25.dim("Tip: Use --force to skip confirmation (required for non-interactive/CI usage).")}
     `).action(async (id, options) => {
     const isJson = getJsonFlag(agents);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -7488,14 +7849,16 @@ ${chalk24.dim("Tip: Use --force to skip confirmation (required for non-interacti
 
 // src/commands/agents/index.ts
 function createAgentsCommand() {
-  const agents = new Command10("agents").description("Manage AI agents").addHelpText("after", `
+  const agents = new Command10("agents").description("Manage AI agents (CRUD only — optimization & evaluation are WIP)").addHelpText("after", `
+${AGENTS_WIP_BANNER}
+
 Examples:
-  ${chalk25.dim("$")} mutagent agents list
-  ${chalk25.dim("$")} mutagent agents get <agent-id>
-  ${chalk25.dim("$")} mutagent agents create --name "Code Reviewer" --slug code-reviewer --system-prompt "You are a code reviewer..."
-  ${chalk25.dim("$")} mutagent agents create -d '{"name":"Code Reviewer","slug":"code-reviewer","systemPrompt":"You are..."}'
-  ${chalk25.dim("$")} mutagent agents update <agent-id> --name "Updated Name"
-  ${chalk25.dim("$")} mutagent agents delete <agent-id> --force
+  ${chalk26.dim("$")} mutagent agents list
+  ${chalk26.dim("$")} mutagent agents get <agent-id>
+  ${chalk26.dim("$")} mutagent agents create --name "Code Reviewer" --slug code-reviewer --system-prompt "You are a code reviewer..."
+  ${chalk26.dim("$")} mutagent agents create -d '{"name":"Code Reviewer","slug":"code-reviewer","systemPrompt":"You are..."}'
+  ${chalk26.dim("$")} mutagent agents update <agent-id> --name "Updated Name"
+  ${chalk26.dim("$")} mutagent agents delete <agent-id> --force
 
 Subcommands:
   list, get, create, update, delete
@@ -7507,22 +7870,22 @@ Subcommands:
 // src/commands/config.ts
 init_config();
 import { Command as Command11 } from "commander";
-import chalk26 from "chalk";
+import chalk27 from "chalk";
 init_errors();
 init_sdk_client();
 var VALID_CONFIG_KEYS = ["apiKey", "endpoint", "format", "timeout", "defaultWorkspace", "defaultOrganization"];
 function createConfigCommand() {
   const config = new Command11("config").description("Manage CLI configuration").addHelpText("after", `
 Examples:
-  ${chalk26.dim("$")} mutagent config list
-  ${chalk26.dim("$")} mutagent config get endpoint
-  ${chalk26.dim("$")} mutagent config set workspace <workspace-id>
-  ${chalk26.dim("$")} mutagent config set org <org-id>
+  ${chalk27.dim("$")} mutagent config list
+  ${chalk27.dim("$")} mutagent config get endpoint
+  ${chalk27.dim("$")} mutagent config set workspace <workspace-id>
+  ${chalk27.dim("$")} mutagent config set org <org-id>
     `);
   config.command("list").description("List all configuration").addHelpText("after", `
 Examples:
-  ${chalk26.dim("$")} mutagent config list
-  ${chalk26.dim("$")} mutagent config list --json
+  ${chalk27.dim("$")} mutagent config list
+  ${chalk27.dim("$")} mutagent config list --json
     `).action(() => {
     const isJson = getJsonFlag(config);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -7535,11 +7898,11 @@ Examples:
   });
   config.command("get").description("Get configuration value").argument("<key>", "Configuration key (apiKey, endpoint, format, timeout, defaultWorkspace, defaultOrganization)").addHelpText("after", `
 Examples:
-  ${chalk26.dim("$")} mutagent config get endpoint
-  ${chalk26.dim("$")} mutagent config get defaultWorkspace
-  ${chalk26.dim("$")} mutagent config get apiKey --json
+  ${chalk27.dim("$")} mutagent config get endpoint
+  ${chalk27.dim("$")} mutagent config get defaultWorkspace
+  ${chalk27.dim("$")} mutagent config get apiKey --json
 
-${chalk26.dim("Keys: apiKey, endpoint, format, timeout, defaultWorkspace, defaultOrganization")}
+${chalk27.dim("Keys: apiKey, endpoint, format, timeout, defaultWorkspace, defaultOrganization")}
     `).action((key) => {
     const isJson = getJsonFlag(config);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -7566,9 +7929,9 @@ ${chalk26.dim("Keys: apiKey, endpoint, format, timeout, defaultWorkspace, defaul
   });
   set.command("workspace").description("Set default workspace ID").argument("<id>", "Workspace ID to set as default").addHelpText("after", `
 Examples:
-  ${chalk26.dim("$")} mutagent config set workspace <workspace-id>
+  ${chalk27.dim("$")} mutagent config set workspace <workspace-id>
 
-${chalk26.dim("Persists workspace ID so you don't need to pass headers on every request.")}
+${chalk27.dim("Persists workspace ID so you don't need to pass headers on every request.")}
     `).action((id) => {
     const isJson = getJsonFlag(config);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -7582,9 +7945,9 @@ ${chalk26.dim("Persists workspace ID so you don't need to pass headers on every
   });
   set.command("org").description("Set default organization ID").argument("<id>", "Organization ID to set as default").addHelpText("after", `
 Examples:
-  ${chalk26.dim("$")} mutagent config set org <org-id>
+  ${chalk27.dim("$")} mutagent config set org <org-id>
 
-${chalk26.dim("Persists organization ID for org-scoped API keys.")}
+${chalk27.dim("Persists organization ID for org-scoped API keys.")}
     `).action((id) => {
     const isJson = getJsonFlag(config);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -7603,7 +7966,7 @@ ${chalk26.dim("Persists organization ID for org-scoped API keys.")}
 // src/commands/playground.ts
 init_sdk_client();
 import { Command as Command12 } from "commander";
-import chalk27 from "chalk";
+import chalk28 from "chalk";
 init_errors();
 function parseSSELine(line3) {
   if (!line3 || line3.startsWith(":")) {
@@ -7630,11 +7993,11 @@ function parsePromptStreamEvent(data) {
 function createPlaygroundCommand() {
   const playground = new Command12("playground").description("Execute and test prompts interactively").addHelpText("after", `
 Examples:
-  ${chalk27.dim("$")} mutagent playground run <prompt-id> --input '{"name": "John"}'
-  ${chalk27.dim("$")} mutagent playground run <prompt-id> --input '{}' --stream
-  ${chalk27.dim("$")} mutagent playground run <prompt-id> -i '{}' --model gpt-4-turbo
-  ${chalk27.dim("$")} mutagent playground run <prompt-id> --system "You are helpful" --human "Hello"
-  ${chalk27.dim("$")} mutagent playground run <prompt-id> --messages '[{"role":"user","content":"Hi"}]'
+  ${chalk28.dim("$")} mutagent playground run <prompt-id> --input '{"name": "John"}'
+  ${chalk28.dim("$")} mutagent playground run <prompt-id> --input '{}' --stream
+  ${chalk28.dim("$")} mutagent playground run <prompt-id> -i '{}' --model gpt-4-turbo
+  ${chalk28.dim("$")} mutagent playground run <prompt-id> --system "You are helpful" --human "Hello"
+  ${chalk28.dim("$")} mutagent playground run <prompt-id> --messages '[{"role":"user","content":"Hi"}]'
 
 Input Format:
   The input must be a valid JSON object matching the prompt's input schema.
@@ -7650,16 +8013,16 @@ Streaming:
     `);
   playground.command("run").description("Execute a prompt with input variables").argument("<prompt-id>", "Prompt ID to execute (from: mutagent prompts list)").option("-i, --input <json>", "Input variables as JSON").option("-s, --stream", "Stream the response").option("-m, --model <model>", "Override model").option("--system <text>", "Set system prompt text").option("--human <text>", "Set human/user message text").option("--messages <json>", "Pass full messages array as JSON string").addHelpText("after", `
 Examples:
-  ${chalk27.dim("$")} mutagent playground run <prompt-id> --input '{"name": "John"}'
-  ${chalk27.dim("$")} mutagent playground run <prompt-id> --input '{}' --stream
-  ${chalk27.dim("$")} mutagent playground run <prompt-id> --system "You are helpful" --human "Hello"
-  ${chalk27.dim("$")} mutagent playground run <prompt-id> --input '{}' --model gpt-4-turbo --json
+  ${chalk28.dim("$")} mutagent playground run <prompt-id> --input '{"name": "John"}'
+  ${chalk28.dim("$")} mutagent playground run <prompt-id> --input '{}' --stream
+  ${chalk28.dim("$")} mutagent playground run <prompt-id> --system "You are helpful" --human "Hello"
+  ${chalk28.dim("$")} mutagent playground run <prompt-id> --input '{}' --model gpt-4-turbo --json
 
 Input Methods (pick one, priority order):
-  --system/--human            Quick system + user message ${chalk27.green("(recommended)")}
+  --system/--human            Quick system + user message ${chalk28.green("(recommended)")}
   --input '{"key":"value"}'   Inline JSON variables
   --messages '[...]'          Full messages array
-${chalk27.dim(`Hint: Test before evaluating: mutagent playground run <id> --input '{"key":"value"}'`)}
+${chalk28.dim(`Hint: Test before evaluating: mutagent playground run <id> --input '{"key":"value"}'`)}
     `).action(async (promptId, options) => {
     const isJson = getJsonFlag(playground);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -7681,21 +8044,21 @@ ${chalk27.dim(`Hint: Test before evaluating: mutagent playground run <id> --inpu
             }
           });
         } else {
-          console.log(chalk27.bold(`
+          console.log(chalk28.bold(`
 Execution Result:`));
-          console.log(chalk27.gray("─".repeat(50)));
-          console.log(chalk27.cyan("Output:"));
+          console.log(chalk28.gray("─".repeat(50)));
+          console.log(chalk28.cyan("Output:"));
           console.log(result.output);
-          console.log(chalk27.gray("─".repeat(50)));
-          console.log(chalk27.dim(`Model: ${result.model}`));
-          console.log(chalk27.dim(`Execution Time: ${String(result.executionTimeMs)}ms`));
+          console.log(chalk28.gray("─".repeat(50)));
+          console.log(chalk28.dim(`Model: ${result.model}`));
+          console.log(chalk28.dim(`Execution Time: ${String(result.executionTimeMs)}ms`));
           if (result.tokens) {
-            console.log(chalk27.dim(`Tokens: ${String(result.tokens.prompt)} prompt + ${String(result.tokens.completion)} completion = ${String(result.tokens.total)} total`));
+            console.log(chalk28.dim(`Tokens: ${String(result.tokens.prompt)} prompt + ${String(result.tokens.completion)} completion = ${String(result.tokens.total)} total`));
           }
           if (result.cost !== undefined) {
-            console.log(chalk27.dim(`Cost: $${result.cost.toFixed(6)}`));
+            console.log(chalk28.dim(`Cost: $${result.cost.toFixed(6)}`));
           }
-          console.log(chalk27.dim(`Playground: ${playgroundLink(promptId)}`));
+          console.log(chalk28.dim(`Playground: ${playgroundLink(promptId)}`));
           console.log();
         }
       }
@@ -7784,9 +8147,9 @@ async function executeStreaming(client, promptId, input, model, isJson, output)
   const decoder = new TextDecoder;
   let buffer = "";
   if (!isJson) {
-    console.log(chalk27.bold(`
+    console.log(chalk28.bold(`
 Streaming Output:`));
-    console.log(chalk27.gray("─".repeat(50)));
+    console.log(chalk28.gray("─".repeat(50)));
   }
   try {
     for (;; ) {
@@ -7825,15 +8188,15 @@ Streaming Output:`));
                 console.log(JSON.stringify({ type: "complete", result: event.result }));
               } else {
                 console.log();
-                console.log(chalk27.gray("─".repeat(50)));
+                console.log(chalk28.gray("─".repeat(50)));
                 if (event.result) {
-                  console.log(chalk27.dim(`Model: ${event.result.model}`));
-                  console.log(chalk27.dim(`Execution Time: ${String(event.result.executionTimeMs)}ms`));
+                  console.log(chalk28.dim(`Model: ${event.result.model}`));
+                  console.log(chalk28.dim(`Execution Time: ${String(event.result.executionTimeMs)}ms`));
                   if (event.result.tokens) {
-                    console.log(chalk27.dim(`Tokens: ${String(event.result.tokens.prompt)} prompt + ${String(event.result.tokens.completion)} completion = ${String(event.result.tokens.total)} total`));
+                    console.log(chalk28.dim(`Tokens: ${String(event.result.tokens.prompt)} prompt + ${String(event.result.tokens.completion)} completion = ${String(event.result.tokens.total)} total`));
                   }
                   if (event.result.cost !== undefined) {
-                    console.log(chalk27.dim(`Cost: $${event.result.cost.toFixed(6)}`));
+                    console.log(chalk28.dim(`Cost: $${event.result.cost.toFixed(6)}`));
                   }
                 }
                 console.log();
@@ -7858,13 +8221,13 @@ Streaming Output:`));
 // src/commands/workspaces.ts
 init_sdk_client();
 import { Command as Command13 } from "commander";
-import chalk28 from "chalk";
+import chalk29 from "chalk";
 init_errors();
 function createWorkspacesCommand() {
   const workspaces = new Command13("workspaces").description("View workspaces (read-only)").addHelpText("after", `
 Examples:
-  ${chalk28.dim("$")} mutagent workspaces list
-  ${chalk28.dim("$")} mutagent workspaces get <workspace-id>
+  ${chalk29.dim("$")} mutagent workspaces list
+  ${chalk29.dim("$")} mutagent workspaces get <workspace-id>
 
 Subcommands:
   list, get
@@ -7873,8 +8236,8 @@ Note: Workspace management (create, update, delete) is available in the Admin Pa
     `);
   workspaces.command("list").description("List all workspaces").option("-l, --limit <n>", "Limit results", "50").option("-o, --offset <n>", "Offset for pagination").addHelpText("after", `
 Examples:
-  ${chalk28.dim("$")} mutagent workspaces list
-  ${chalk28.dim("$")} mutagent workspaces list --limit 10 --json
+  ${chalk29.dim("$")} mutagent workspaces list
+  ${chalk29.dim("$")} mutagent workspaces list --limit 10 --json
     `).action(async (options) => {
     const isJson = getJsonFlag(workspaces);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -7914,8 +8277,8 @@ Examples:
   });
   workspaces.command("get").description("Get workspace details").argument("<id>", "Workspace ID").addHelpText("after", `
 Examples:
-  ${chalk28.dim("$")} mutagent workspaces get <workspace-id>
-  ${chalk28.dim("$")} mutagent workspaces get <workspace-id> --json
+  ${chalk29.dim("$")} mutagent workspaces get <workspace-id>
+  ${chalk29.dim("$")} mutagent workspaces get <workspace-id> --json
     `).action(async (id) => {
     const isJson = getJsonFlag(workspaces);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -7948,12 +8311,12 @@ Examples:
 // src/commands/providers/index.ts
 init_sdk_client();
 import { Command as Command14 } from "commander";
-import chalk32 from "chalk";
+import chalk33 from "chalk";
 init_errors();
 
 // src/commands/providers/add.ts
 init_sdk_client();
-import chalk29 from "chalk";
+import chalk30 from "chalk";
 init_errors();
 function resolveScope(scopeFlag, client) {
   const scope = scopeFlag ?? "workspace";
@@ -7976,10 +8339,10 @@ function resolveScope(scopeFlag, client) {
 function registerAddCommand(parent) {
   parent.command("add").description("Add a new provider configuration").requiredOption("-p, --provider <type>", "Provider type (openai, anthropic, google, ...)").requiredOption("-n, --name <name>", "Display name for this provider").requiredOption("-k, --api-key <key>", "API key for the provider").option("-s, --scope <scope>", "Scope: workspace (default), org, user", "workspace").option("--base-url <url>", "Custom base URL for the provider API").option("--set-default", "Set as default provider for this scope").addHelpText("after", `
 Examples:
-  ${chalk29.dim("$")} mutagent providers add --provider openai --name "My OpenAI" --api-key $OPENAI_API_KEY
-  ${chalk29.dim("$")} mutagent providers add --provider anthropic --name "Team Claude" --api-key $KEY --scope org
-  ${chalk29.dim("$")} mutagent providers add --provider openai --name "Personal" --api-key $KEY --scope user --set-default
-  ${chalk29.dim("$")} mutagent providers add --provider custom --name "Ollama" --api-key none --base-url http://localhost:11434 --json
+  ${chalk30.dim("$")} mutagent providers add --provider openai --name "My OpenAI" --api-key $OPENAI_API_KEY
+  ${chalk30.dim("$")} mutagent providers add --provider anthropic --name "Team Claude" --api-key $KEY --scope org
+  ${chalk30.dim("$")} mutagent providers add --provider openai --name "Personal" --api-key $KEY --scope user --set-default
+  ${chalk30.dim("$")} mutagent providers add --provider custom --name "Ollama" --api-key none --base-url http://localhost:11434 --json
 
 Scope Resolution:
   workspace (default)   Uses your configured workspace
@@ -8025,10 +8388,10 @@ The API key is encrypted server-side and never returned in plain text.
         console.log(`  Scope: ${options.scope ?? "workspace"}`);
         console.log(`  URL:   ${providerLink(created.id)}`);
         if (options.setDefault) {
-          console.log(chalk29.green("  Set as default provider"));
+          console.log(chalk30.green("  Set as default provider"));
         }
         console.log("");
-        console.log(chalk29.dim("API key is encrypted server-side. Use `providers get` to see masked key."));
+        console.log(chalk30.dim("API key is encrypted server-side. Use `providers get` to see masked key."));
       }
     } catch (error) {
       handleError(error, isJson);
@@ -8066,15 +8429,15 @@ function buildProviderCreatedDirective(provider, scope) {
 
 // src/commands/providers/update.ts
 init_sdk_client();
-import chalk30 from "chalk";
+import chalk31 from "chalk";
 init_errors();
 function registerUpdateCommand(parent) {
   parent.command("update").description("Update an existing provider configuration").argument("<id>", "Provider ID (from: mutagent providers list)").option("-n, --name <name>", "Updated display name").option("-k, --api-key <key>", "Updated API key (will be re-encrypted)").option("--active <bool>", "Activate or deactivate (true|false)").option("--set-default", "Set as default provider for its scope").option("--base-url <url>", 'Updated base URL (use "" to clear)').addHelpText("after", `
 Examples:
-  ${chalk30.dim("$")} mutagent providers update <id> --name "New Name"
-  ${chalk30.dim("$")} mutagent providers update <id> --api-key $NEW_KEY --json
-  ${chalk30.dim("$")} mutagent providers update <id> --active false
-  ${chalk30.dim("$")} mutagent providers update <id> --set-default --json
+  ${chalk31.dim("$")} mutagent providers update <id> --name "New Name"
+  ${chalk31.dim("$")} mutagent providers update <id> --api-key $NEW_KEY --json
+  ${chalk31.dim("$")} mutagent providers update <id> --active false
+  ${chalk31.dim("$")} mutagent providers update <id> --set-default --json
 
 PATCH semantics — only provided fields are updated.
     `).action(async (id, options) => {
@@ -8125,7 +8488,7 @@ PATCH semantics — only provided fields are updated.
         console.log(`  ID:  ${String(updated.id ?? id)}`);
         console.log(`  URL: ${providerLink(updated.id ?? id)}`);
         if (options.apiKey) {
-          console.log(chalk30.dim("  API key re-encrypted server-side."));
+          console.log(chalk31.dim("  API key re-encrypted server-side."));
         }
       }
     } catch (error) {
@@ -8164,17 +8527,17 @@ function buildProviderUpdatedDirective(provider, requestId) {
 
 // src/commands/providers/delete.ts
 init_sdk_client();
-import chalk31 from "chalk";
+import chalk32 from "chalk";
 init_errors();
 function registerDeleteCommand(parent) {
   parent.command("delete").description("Delete a provider configuration").argument("<id>", "Provider ID (from: mutagent providers list)").option("-f, --force", "Skip confirmation prompt").addHelpText("after", `
 Examples:
-  ${chalk31.dim("$")} mutagent providers delete <id>
-  ${chalk31.dim("$")} mutagent providers delete <id> --force
-  ${chalk31.dim("$")} mutagent providers delete <id> --force --json
+  ${chalk32.dim("$")} mutagent providers delete <id>
+  ${chalk32.dim("$")} mutagent providers delete <id> --force
+  ${chalk32.dim("$")} mutagent providers delete <id> --force --json
 
-${chalk31.dim("Note: --force is required. The CLI is non-interactive — confirm with the user via your native flow, then pass --force.")}
-${chalk31.dim("Warning: API keys are AES-256-GCM encrypted and irrecoverable after deletion. Agents referencing this provider will lose their model config.")}
+${chalk32.dim("Note: --force is required. The CLI is non-interactive — confirm with the user via your native flow, then pass --force.")}
+${chalk32.dim("Warning: API keys are AES-256-GCM encrypted and irrecoverable after deletion. Agents referencing this provider will lose their model config.")}
     `).action(async (id, options) => {
     const isJson = getJsonFlag(parent);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -8264,12 +8627,12 @@ function validateProviderType(type) {
 function createProvidersCommand() {
   const providers = new Command14("providers").description("Manage LLM provider configurations (BYOK)").addHelpText("after", `
 Examples:
-  ${chalk32.dim("$")} mutagent providers list
-  ${chalk32.dim("$")} mutagent providers get <provider-id>
-  ${chalk32.dim("$")} mutagent providers add --provider openai --name "My OpenAI" --api-key $KEY
-  ${chalk32.dim("$")} mutagent providers update <id> --name "New Name"
-  ${chalk32.dim("$")} mutagent providers delete <id> --force
-  ${chalk32.dim("$")} mutagent providers test <provider-id>
+  ${chalk33.dim("$")} mutagent providers list
+  ${chalk33.dim("$")} mutagent providers get <provider-id>
+  ${chalk33.dim("$")} mutagent providers add --provider openai --name "My OpenAI" --api-key $KEY
+  ${chalk33.dim("$")} mutagent providers update <id> --name "New Name"
+  ${chalk33.dim("$")} mutagent providers delete <id> --force
+  ${chalk33.dim("$")} mutagent providers test <provider-id>
 
 Provider Types:
   openai, anthropic, google, azure, bedrock, cohere, mistral, groq, together, replicate, custom
@@ -8279,9 +8642,9 @@ Subcommands:
     `);
   providers.command("list").description("List all providers").option("-l, --limit <n>", "Limit results", "50").option("-o, --offset <n>", "Offset for pagination").option("-t, --type <type>", "Filter by provider type").addHelpText("after", `
 Examples:
-  ${chalk32.dim("$")} mutagent providers list
-  ${chalk32.dim("$")} mutagent providers list --type openai
-  ${chalk32.dim("$")} mutagent providers list --json
+  ${chalk33.dim("$")} mutagent providers list
+  ${chalk33.dim("$")} mutagent providers list --type openai
+  ${chalk33.dim("$")} mutagent providers list --json
     `).action(async (options) => {
     const isJson = getJsonFlag(providers);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -8330,8 +8693,8 @@ Examples:
   });
   providers.command("get").description("Get provider details").argument("<id>", "Provider ID (from: mutagent providers list)").addHelpText("after", `
 Examples:
-  ${chalk32.dim("$")} mutagent providers get <provider-id>
-  ${chalk32.dim("$")} mutagent providers get <provider-id> --json
+  ${chalk33.dim("$")} mutagent providers get <provider-id>
+  ${chalk33.dim("$")} mutagent providers get <provider-id> --json
     `).action(async (id) => {
     const isJson = getJsonFlag(providers);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -8360,10 +8723,10 @@ Examples:
   });
   providers.command("test").description("Test provider connectivity").argument("<id>", "Provider ID (from: mutagent providers list)").addHelpText("after", `
 Examples:
-  ${chalk32.dim("$")} mutagent providers test <provider-id>
-  ${chalk32.dim("$")} mutagent providers test <provider-id> --json
+  ${chalk33.dim("$")} mutagent providers test <provider-id>
+  ${chalk33.dim("$")} mutagent providers test <provider-id> --json
 
-${chalk32.dim("Tests connectivity and lists available models for the provider.")}
+${chalk33.dim("Tests connectivity and lists available models for the provider.")}
     `).action(async (id) => {
     const isJson = getJsonFlag(providers);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -8378,9 +8741,9 @@ ${chalk32.dim("Tests connectivity and lists available models for the provider.")
       } else {
         if (result.success) {
           output.success(`Provider test passed (${String(result.responseTimeMs)}ms)`);
-          console.log(chalk32.green(`Message: ${result.message}`));
+          console.log(chalk33.green(`Message: ${result.message}`));
           if (result.availableModels && result.availableModels.length > 0) {
-            console.log(chalk32.bold(`
+            console.log(chalk33.bold(`
 Available Models:`));
             result.availableModels.forEach((model) => {
               console.log(`  - ${model}`);
@@ -8389,7 +8752,7 @@ Available Models:`));
         } else {
           output.error(`Provider test failed: ${result.message}`);
           if (result.error) {
-            console.log(chalk32.red(`Error: ${result.error}`));
+            console.log(chalk33.red(`Error: ${result.error}`));
           }
         }
       }
@@ -8407,7 +8770,7 @@ Available Models:`));
 init_config();
 import { Command as Command15 } from "commander";
 import inquirer2 from "inquirer";
-import chalk33 from "chalk";
+import chalk34 from "chalk";
 import { existsSync as existsSync11, mkdirSync as mkdirSync3, writeFileSync as writeFileSync4 } from "fs";
 import { execSync as execSync3 } from "child_process";
 import { join as join6 } from "path";
@@ -8530,13 +8893,13 @@ function writeRcConfig(config, cwd = process.cwd()) {
 function createInitCommand() {
   const init = new Command15("init").description("Initialize MutagenT in your project").option("--non-interactive", "Skip interactive prompts (defaults to CLI-only mode)").addHelpText("after", `
 Examples:
-  ${chalk33.dim("$")} mutagent init                     # Interactive setup wizard
-  ${chalk33.dim("$")} mutagent init --non-interactive    # CLI-only mode (no prompts)
+  ${chalk34.dim("$")} mutagent init                     # Interactive setup wizard
+  ${chalk34.dim("$")} mutagent init --non-interactive    # CLI-only mode (no prompts)
 
 Modes:
-  ${chalk33.bold("Full scaffold")}     Install SDK + integration package, create config, setup tracing
-  ${chalk33.bold("CLI-only")}          Verify auth + create .mutagentrc.json with workspace/endpoint
-  ${chalk33.bold("Skip")}              Exit without changes
+  ${chalk34.bold("Full scaffold")}     Install SDK + integration package, create config, setup tracing
+  ${chalk34.bold("CLI-only")}          Verify auth + create .mutagentrc.json with workspace/endpoint
+  ${chalk34.bold("Skip")}              Exit without changes
     `).action(async (options) => {
     const isJson = getJsonFlag(init);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -8808,23 +9171,36 @@ Modes:
 
 // src/commands/explore.ts
 import { Command as Command16 } from "commander";
-import chalk34 from "chalk";
+import chalk35 from "chalk";
 import { resolve as resolve3 } from "path";
 init_errors();
 function createExploreCommand() {
   const explore = new Command16("explore").description("Scan codebase for prompts, datasets, and MutagenT markers").option("-p, --path <dir>", "Directory to scan", ".").option("--depth <n>", "Max directory depth", "10").option("--include <glob>", "Include file pattern", "**/*.{ts,js,py,tsx,jsx}").option("--exclude <dirs>", "Comma-separated directories to exclude", "node_modules,dist,.git,build,.next,__pycache__,venv,.venv").option("--markers-only", "Only find existing MutagenT markers").addHelpText("after", `
 Examples:
-  ${chalk34.dim("$")} mutagent explore
-  ${chalk34.dim("$")} mutagent explore --path ./src
-  ${chalk34.dim("$")} mutagent explore --include "**/*.{ts,py}" --depth 5
-  ${chalk34.dim("$")} mutagent explore --markers-only
-  ${chalk34.dim("$")} mutagent explore --json
+  ${chalk35.dim("$")} mutagent explore
+  ${chalk35.dim("$")} mutagent explore --path ./src
+  ${chalk35.dim("$")} mutagent explore --include "**/*.{ts,py}" --depth 5
+  ${chalk35.dim("$")} mutagent explore --markers-only
+  ${chalk35.dim("$")} mutagent explore --json
 
 Detection modes:
-  ${chalk34.dim("Heuristic")}  Template variables ({{var}}), prompt constants, schema definitions
-  ${chalk34.dim("Marker")}     MutagenT:START/END comment markers from previous uploads
+  ${chalk35.dim("Heuristic")}  Template variables ({var} or {{var}}), prompt constants, schema definitions
+  ${chalk35.dim("Marker")}     MutagenT:START/END comment markers from previous uploads
+  ${chalk35.dim("Agents")}     Multi-turn / tool-calling patterns (informational — WIP)
+
+${chalk35.bold("Template variable delimiters (inferred per-file):")}
+  ${chalk35.dim("single")}    ${chalk35.cyan("{variable}")}    MutagenT platform canonical; LangChain PromptTemplate
+  ${chalk35.dim("double")}    ${chalk35.cyan("{{variable}}")}  Handlebars / Mustache / LiquidJS / Jinja2 / LangChain ChatPromptTemplate
+
+  ${chalk35.dim("Inference is per-file: majority of bracket style wins, tie goes to single")}
+  ${chalk35.dim('(platform canonical). Fenced code blocks are stripped first; {"json": ...} ')}
+  ${chalk35.dim("literals are skipped via JSON-key lookahead. See the mutagent-cli skill at ")}
+  ${chalk35.dim("concepts/prompt-variables.md for the full contract.")}
 
-${chalk34.dim("Results are saved to .mutagent/mutation-context.md for use by other commands.")}
+  ${chalk35.dim("With --json, each discovered prompt carries a ")}${chalk35.cyan("delimiter")}${chalk35.dim(" field:")}
+    ${chalk35.dim('{ "file": "src/prompts/summarize.ts", "line": 12, "reason": "template-variable", "delimiter": "single" }')}
+
+${chalk35.dim("Results are saved to .mutagent/mutation-context.md for use by other commands.")}
     `).action((options) => {
     const isJson = getJsonFlag(explore);
     const output = new OutputFormatter(isJson ? "json" : "table");
@@ -8842,7 +9218,7 @@ ${chalk34.dim("Results are saved to .mutagent/mutation-context.md for use by oth
         markersOnly
       };
       if (!isJson) {
-        console.log(chalk34.cyan(`
+        console.log(chalk35.cyan(`
 Scanning ${scanPath}...
 `));
       }
@@ -8860,52 +9236,78 @@ Scanning ${scanPath}...
           prompts: result.prompts,
           datasets: result.datasets,
           markers: result.markers,
+          agents: result.agents,
+          agentsWip: result.agents.length > 0 ? {
+            status: "early-access",
+            message: "Agent detection is informational only. Agent optimization & evaluation are WIP.",
+            earlyAccessUrl: "https://www.mutagent.io/agents-partnership"
+          } : undefined,
           summary: {
             totalPrompts: result.prompts.length,
             totalDatasets: result.datasets.length,
-            totalMarkers: result.markers.length
+            totalMarkers: result.markers.length,
+            totalAgents: result.agents.length
           },
           contextFile: ".mutagent/mutation-context.md"
         });
       } else {
-        const totalFindings = result.prompts.length + result.datasets.length + result.markers.length;
+        const totalFindings = result.prompts.length + result.datasets.length + result.markers.length + result.agents.length;
         if (totalFindings === 0) {
           output.info("No prompts, datasets, or markers found.");
-          console.log(chalk34.dim(`
+          console.log(chalk35.dim(`
   Tip: Create a prompt with template variables like {{input}} to get started.`));
           return;
         }
         if (result.prompts.length > 0) {
-          console.log(chalk34.bold(`  Prompts Found (${String(result.prompts.length)}):`));
+          console.log(chalk35.bold(`  Prompts Found (${String(result.prompts.length)}):`));
           console.log();
           for (const p of result.prompts) {
-            const confidenceTag = p.confidence === "high" ? chalk34.green("[high]") : p.confidence === "medium" ? chalk34.yellow("[medium]") : chalk34.dim("[low]");
-            const reasonTag = chalk34.dim(`[${p.reason}]`);
-            console.log(`    ${confidenceTag} ${chalk34.green(p.file)}:${chalk34.yellow(String(p.line))} ${reasonTag}`);
-            console.log(`      ${chalk34.dim(p.preview)}`);
+            const confidenceTag = p.confidence === "high" ? chalk35.green("[high]") : p.confidence === "medium" ? chalk35.yellow("[medium]") : chalk35.dim("[low]");
+            const reasonTag = chalk35.dim(`[${p.reason}]`);
+            console.log(`    ${confidenceTag} ${chalk35.green(p.file)}:${chalk35.yellow(String(p.line))} ${reasonTag}`);
+            console.log(`      ${chalk35.dim(p.preview)}`);
           }
           console.log();
         }
         if (result.datasets.length > 0) {
-          console.log(chalk34.bold(`  Datasets Found (${String(result.datasets.length)}):`));
+          console.log(chalk35.bold(`  Datasets Found (${String(result.datasets.length)}):`));
           console.log();
           for (const d of result.datasets) {
-            console.log(`    ${chalk34.green(d.file)} ${chalk34.dim(`(${String(d.items)} items)`)}`);
+            console.log(`    ${chalk35.green(d.file)} ${chalk35.dim(`(${String(d.items)} items)`)}`);
           }
           console.log();
         }
         if (result.markers.length > 0) {
-          console.log(chalk34.bold(`  MutagenT Markers (${String(result.markers.length)}):`));
+          console.log(chalk35.bold(`  MutagenT Markers (${String(result.markers.length)}):`));
           console.log();
           for (const m of result.markers) {
-            const idPart = m.platformId ? chalk34.cyan(` id=${m.platformId}`) : "";
-            console.log(`    ${chalk34.green(m.file)}:${chalk34.yellow(String(m.line))} ${chalk34.magenta(m.type)}${idPart}`);
+            const idPart = m.platformId ? chalk35.cyan(` id=${m.platformId}`) : "";
+            console.log(`    ${chalk35.green(m.file)}:${chalk35.yellow(String(m.line))} ${chalk35.magenta(m.type)}${idPart}`);
+          }
+          console.log();
+        }
+        if (result.agents.length > 0) {
+          console.log(chalk35.bold(`  Agents Detected (${String(result.agents.length)}):`));
+          console.log();
+          for (const a of result.agents) {
+            const confidenceTag = a.confidence === "high" ? chalk35.green("[high]") : a.confidence === "medium" ? chalk35.yellow("[medium]") : chalk35.dim("[low]");
+            const reasonTag = chalk35.dim(`[${a.reason}]`);
+            console.log(`    ${confidenceTag} ${chalk35.green(a.file)}:${chalk35.yellow(String(a.line))} ${reasonTag}`);
+            console.log(`      ${chalk35.dim(a.preview)}`);
           }
           console.log();
+          console.log(chalk35.yellow("  ┌─ Agents: Early Access / WIP ──────────────────────────────"));
+          console.log(chalk35.yellow("  │ ") + chalk35.bold("Agent optimization & evaluation are actively in development."));
+          console.log(chalk35.yellow("  │ ") + chalk35.dim("Detected patterns above are informational only."));
+          console.log(chalk35.yellow("  │ ") + chalk35.dim("The Prompt path (single-shot + output schema) is production-ready."));
+          console.log(chalk35.yellow("  │ ") + chalk35.dim("Join early access for the Agent path:"));
+          console.log(chalk35.yellow("  │ ") + chalk35.cyan("https://www.mutagent.io/agents-partnership"));
+          console.log(chalk35.yellow("  └────────────────────────────────────────────────────────────"));
+          console.log();
         }
-        console.log(chalk34.dim("  ─────────────────────────────────"));
-        console.log(`  ${chalk34.bold("Summary:")} ${String(result.prompts.length)} prompts, ${String(result.datasets.length)} datasets, ${String(result.markers.length)} markers`);
-        console.log(chalk34.dim(`  Saved to .mutagent/mutation-context.md`));
+        console.log(chalk35.dim("  ─────────────────────────────────"));
+        console.log(`  ${chalk35.bold("Summary:")} ${String(result.prompts.length)} prompts, ${String(result.datasets.length)} datasets, ${String(result.markers.length)} markers, ${String(result.agents.length)} agents`);
+        console.log(chalk35.dim(`  Saved to .mutagent/mutation-context.md`));
         console.log();
       }
     } catch (error) {
@@ -8917,181 +9319,1180 @@ Scanning ${scanPath}...
 
 // src/commands/skills.ts
 import { Command as Command17 } from "commander";
-import chalk35 from "chalk";
+import chalk36 from "chalk";
 import { existsSync as existsSync12, mkdirSync as mkdirSync4, writeFileSync as writeFileSync5 } from "fs";
-import { join as join7 } from "path";
+import { dirname, join as join7 } from "path";
 import { execSync as execSync4 } from "child_process";
-function findRepoRoot() {
-  try {
-    return execSync4("git rev-parse --show-toplevel", { encoding: "utf-8", stdio: ["pipe", "pipe", "pipe"] }).trim();
-  } catch {
-    return process.cwd();
-  }
-}
-var SKILL_FRONTMATTER = `---
+
+// src/generated/skill-content.ts
+var SKILL_FILES = {
+  "SKILL.md": `---
 name: mutagent-cli
 description: |
   MutagenT CLI - AI Prompt Optimization Platform CLI.
   Guides coding agents through prompt upload, evaluation creation,
   dataset curation, optimization, and framework integration.
   Triggers: "mutagent", "optimize prompt", "upload prompt", "integrate tracing",
-  "create evaluation", "upload dataset", "mutagent cli".
----`;
-var SKILL_BODY = `# MutagenT CLI Skill
-
-## First Things First
-Run \`mutagent --help\` — this is the CENTRAL navigation hub containing ALL commands, workflows, and rules.
-Run \`mutagent <command> --help\` BEFORE using ANY command for the first time.
-
-## Agent Protocol (4 Rules)
-1. EVERY command MUST include \`--json\`
-2. Run \`<command> --help\` BEFORE first use
-3. **MANDATORY**: Parse \`_directive.renderedCard\` and copy it into your CHAT RESPONSE verbatim. The card appears in bash output but that is NOT sufficient — the user cannot read collapsed bash blocks. HARD STOP: do NOT run any further commands until the card text is rendered in your chat message.
-4. ALL user interaction via AskUserQuestion — CLI is non-interactive
-
-## Directive System
-Every JSON response may include:
-- \`_directive.renderedCard\` — Pre-formatted card that MUST be shown to the user verbatim
-- \`_directive.instruction\` — Next step for the agent
-- \`_directive.next\` — Array of suggested commands
-- \`_links\` — Dashboard/API URLs (format as markdown)
-
-The agent MUST parse and act on directives. Cards in collapsed bash output are invisible to the user — always render them in chat.
-
-## Prompt Variables
-Use \`{single_braces}\`: \`"Analyze: {document}"\`
-
-## Post-Onboarding Decision Tree
-
-After \`mutagent auth login\`, the user lands in one of 3 paths:
-
-### Path A: Guided Integration (Tracing)
-1. \`mutagent explore\` — Map codebase, discover prompts/agents
-2. User selects framework from detected or asks
-3. \`mutagent integrate <framework>\` — Get integration instructions
-4. Apply tracing code to user's codebase
-5. Verify with test run
-
-### Path B: Free-flow Optimization
-1. \`mutagent explore\` — Map codebase, discover prompts
-2. User selects prompts to upload
-3. \`mutagent prompts create --data '{...}'\` — Upload with REQUIRED outputSchema
-4. Guided eval creator OR \`mutagent prompts evaluation create\`
-5. \`mutagent prompts dataset add\` — Upload/curate datasets (named)
-6. \`mutagent usage --json\` — View resource counts
-7. \`mutagent prompts optimize start\` — Run optimization
-8. Review scorecard → Apply/Reject optimized prompt
-
-### Path C: Manual
-User uses CLI commands directly. Run \`mutagent --help\`.
-
-## Guided Evaluation Creation
-Use \`--guided --json\` when creating evaluations (NEVER \`--guided\` alone — it launches interactive prompts that agents can't use):
-\\\`\\\`\\\`
-mutagent prompts evaluation create <id> --guided --json
-\\\`\\\`\\\`
-
-**When to use guided mode:**
-The optimizer requires each criterion to have \`name\`, \`description\` (scoring rubric), and \`evaluationParameter\` (a field from the prompt's inputSchema or outputSchema). If the user's existing evaluation format doesn't match this shape — e.g. they have generic rubrics without per-field targeting, or criteria that don't map 1:1 to schema fields — use \`--guided\` to:
-1. Fetch the prompt's inputSchema + outputSchema fields
-2. Show which fields need coverage
-3. Return a criteria template with the exact JSON shape
-
-**Validation rules the CLI enforces:**
-- Every criterion MUST have \`name\`, \`description\`, and \`evaluationParameter\`
-- \`evaluationParameter\` MUST match an actual schema field name
-- No duplicate \`evaluationParameter\` values — each criterion targets a unique field
-- ALL schema fields must be covered (missing fields = error)
+  "create evaluation", "upload dataset", "explore prompts", "mutagent cli".
+---
+
+# MutagenT CLI Skill
+
+> **Canonical source**: \`mutagent-cli/.claude/skills/mutagent-cli/SKILL.md\`
+> Packed into the CLI binary via \`scripts/sync-skill.ts\`. Installed to end-user
+> dev environments via \`mutagent skills install\`. Edit this file, not the installed copy.
+
+## SKILL vs CLI — responsibility split
+
+| Layer | Owner | Responsibility |
+|---|---|---|
+| **SKILL** (this file + subfiles) | here | journeys, routing, 5 rules, enforcement |
+| **CLI** | \`mutagent <cmd>\` | commands, flags, \`--json\`, \`_directive.*\`, \`_links\` |
+| Platform | api.mutagent.io | storage, optimization, eval execution, \`{variable}\` rendering |
+
+**Rule**: SKILL never duplicates CLI flag lists — always \`mutagent <cmd> --help\` for flags.
+
+---
+
+## 5 Core Rules — NON-NEGOTIABLE
+
+1. **\`--json\` on EVERY command.** No exceptions. Agents use JSON mode exclusively.
+2. **\`<command> --help\` BEFORE first use of any command.** The CLI is the source of truth for flags — this SKILL never inlines them.
+3. **NEVER auto-generate eval criteria — collect from user.** Ask the user for each rubric field. See [concepts/eval-criteria.md](./concepts/eval-criteria.md) for the 3-tier format.
+4. **Explore-before-modify.** Run \`mutagent explore --json\` before any write operation. Present findings, get user confirmation. Never mutate without discovery first.
+5. **Cost transparency before \`optimize start\`.** Run \`mutagent usage --json\` and show the result to the user. Get explicit confirmation before any optimization job.
+
+---
+
+## Journey Router — route by user intent
+
+Match the user's first request. Load ONLY the matching subfile. Do NOT preload the whole set.
+
+| User said / signal detected | Load subfile | Why |
+|---|---|---|
+| "trace", "observe", "integrate", "add framework" | [workflows/tracing.md](./workflows/tracing.md) | Non-destructive, fastest first-value path |
+| "optimize", "improve", "tune", "evaluate", "upload prompt" | [workflows/optimization.md](./workflows/optimization.md) | Full create→dataset→eval→optimize loop |
+| "explore", "scan", "find prompts", "what prompts", "discover" | [workflows/exploration.md](./workflows/exploration.md) | Read-only discovery + taxonomy |
+| \`AgentExecutor\`, \`StateGraph\`, \`createReactAgent\`, \`tool_calls\`, \`@tool\`, \`langgraph\`, \`crewai\`, \`autogen\`, \`openai/agents\`, multi-turn | [workflows/agents.md](./workflows/agents.md) | WIP path — surface partnership link |
+| "how do variables work", "single vs double braces", delimiter | [concepts/prompt-variables.md](./concepts/prompt-variables.md) | Delimiter inference contract |
+| "what makes a good eval", "rubric", "evaluation criteria" | [concepts/eval-criteria.md](./concepts/eval-criteria.md) | INPUT MVC + OUTPUT Standards |
+| Unclear / first time | run \`mutagent explore --json\` first, then reroute | Discovery before action |
+
+---
+
+## Subfile Map
+
+| File | WHEN to load | WHY | ENFORCEMENT |
+|---|---|---|---|
+| [workflows/tracing.md](./workflows/tracing.md) | User wants to add framework tracing / observability | Non-destructive append-only integration sequence | Must run explore first (Rule 4) |
+| [workflows/optimization.md](./workflows/optimization.md) | User wants to optimize or evaluate a prompt | Full loop: explore → upload → dataset → eval → optimize → apply | Must check usage before optimize (Rule 5); must collect rubrics from user (Rule 3) |
+| [workflows/exploration.md](./workflows/exploration.md) | User wants to scan codebase, identify prompts vs agents | Read-only discovery; output taxonomy to user | Run only; no writes |
+| [workflows/agents.md](./workflows/agents.md) | Multi-turn / tool-calling code detected | WIP — do NOT attempt optimizer, surface partnership link | Show WIP card to user verbatim |
+| [concepts/prompt-variables.md](./concepts/prompt-variables.md) | Any question about \`{var}\` vs \`{{var}}\`, delimiter inference | Brace convention + conversion rules | Load before \`prompts create\` in optimization workflow |
+| [concepts/eval-criteria.md](./concepts/eval-criteria.md) | Any question about rubric design, MVC, Output Standards | 3-tier rubric format — INPUT-param vs OUTPUT-param scope | Load before \`evaluation create --guided\` in optimization workflow |
+
+---
+
+## Output handling
+
+After every CLI command:
+- **Show the command output to the user.** Command output appears in bash blocks that users may not see — always present the key results in your chat response.
+- **For evaluation create \`--guided\`**: the CLI provides per-field questions for each schema field. Ask the user each question in turn. Do not skip any field. Do not pre-fill answers.
+- **For \`optimize results\`**: present the before/after scorecard to the user and confirm whether to apply, view diff, or reject.
+
+---
+
+## Anti-patterns — NEVER do these
+
+- Run any command without \`--json\`
+- Auto-generate eval criteria — always collect from the user
+- Skip any schema field when collecting evaluation rubrics
+- Skip \`mutagent explore --json\` before any write operation
+- Run \`optimize start\` without first showing \`usage --json\` to the user
+- Increase \`--max-iterations\` above 1 without explicit user consent (each iteration = LLM spend)
+- Run a multi-turn agent through the prompt optimizer
+- Skip showing command output results to the user
+- Inline CLI flags from memory — always read \`--help\` first
+
+---
 
 ## State Tracking
-- \`.mutagent/mutation-context.md\` — Codebase index of discovered/uploaded prompts
-- Check and update this file after explore, create, and dataset operations
-- \`mutagent auth status\` — Auth + workspace state
-- Comment markers (\`// MutagenT:START ... // MutagenT:END\`) in source files
-
-## Post-Optimization
-After \`optimize results\`: ALWAYS show the before/after diff to the user first. Then offer choices via AskUserQuestion: Apply / Reject.
-
-## Evaluation Criteria Reminder
-Every evaluation MUST specify criteria targeting either:
-- Input variable fields (from inputSchema)
-- Output fields (from outputSchema / structured output)
-
-## Optimization Cost Control
-- Default max-iterations is 1. NEVER increase without explicit user request.
-- Each iteration incurs LLM costs. Starting >1 iterations without consent = protocol violation.
-- If user wants more, confirm the number via AskUserQuestion first.
-
-## Claude Code Hooks (Session Telemetry)
-Optional — sends session activity to MutagenT traces API for observability.
-
-Install via \`.claude/settings.local.json\`:
-\\\`\\\`\\\`json
-{"hooks": {
-  "SessionStart": [{"matcher": "startup", "hooks": [{"type": "command", "command": "mutagent hooks claude-code session-start"}]}],
-  "Stop": [{"hooks": [{"type": "command", "command": "mutagent hooks claude-code session-end"}]}],
-  "PreToolUse": [{"hooks": [{"type": "command", "command": "mutagent hooks claude-code pre-tool-use"}]}],
-  "PostToolUse": [{"hooks": [{"type": "command", "command": "mutagent hooks claude-code post-tool-use"}]}]
-}}
-\\\`\\\`\\\`
-Or run: \`mutagent init\` (option 1 installs skill + hooks together)`;
-var SKILL_DIR = ".claude/skills/mutagent-cli";
-var SKILL_FILE = "SKILL.md";
-function createSkillsCommand() {
-  const skills = new Command17("skills").description("Manage MutagenT CLI skills for coding agents");
-  skills.command("install").description("Install MutagenT CLI skill for Claude Code").addHelpText("after", `
-Examples:
-  ${chalk35.dim("$")} mutagent skills install
 
-This creates a Claude Code skill at .claude/skills/mutagent-cli/SKILL.md
-that teaches coding agents how to use the MutagenT CLI effectively.
-    `).action((_options, cmd) => {
-    const parentCmd = cmd.parent?.parent;
-    const isJson = parentCmd ? getJsonFlag(parentCmd) : false;
-    const output = new OutputFormatter(isJson ? "json" : "table");
-    const repoRoot = findRepoRoot();
-    const skillDir = join7(repoRoot, SKILL_DIR);
-    const skillPath = join7(skillDir, SKILL_FILE);
-    if (!existsSync12(skillDir)) {
-      mkdirSync4(skillDir, { recursive: true });
-    }
-    const content = `${SKILL_FRONTMATTER}
+- \`.mutagent/mutation-context.md\` — codebase index of discovered/uploaded prompts. Update after explore, create, dataset ops.
+- \`mutagent auth status --json\` — auth + workspace state.
 
-${SKILL_BODY}
-`;
-    writeFileSync5(skillPath, content, "utf-8");
-    if (isJson) {
-      output.output({
-        installed: true,
-        path: skillPath,
-        name: "mutagent-cli"
-      });
-    } else {
-      output.success(`Installed MutagenT CLI skill`);
-      console.log(`  ${chalk35.dim("Path:")} ${skillPath}`);
-      console.log("");
-      console.log(`  ${chalk35.dim("This skill teaches coding agents how to use the MutagenT CLI.")}`);
-      console.log(`  ${chalk35.dim("It will be automatically loaded by Claude Code when relevant triggers match.")}`);
-    }
-  });
-  return skills;
-}
+---
 
-// src/commands/usage.ts
-init_config();
-import { Command as Command18 } from "commander";
-import chalk36 from "chalk";
-init_errors();
-init_sdk_client();
-var PROVIDERS_URL = "https://app.mutagent.io/settings/providers";
-function createUsageCommand() {
-  const usage = new Command18("usage").description("Show resource counts (prompts, datasets, evaluations, optimizations, experiments)").addHelpText("after", `
-Examples:
-  ${chalk36.dim("$")} mutagent usage
-  ${chalk36.dim("$")} mutagent usage --json
-    `);
-  usage.action(async () => {
-    const isJson = getJsonFlag(usage);
+## Login (two paths)
+
+- **CI / automated**: \`export MUTAGENT_API_KEY=mt_... && mutagent login --json\` — no browser, no prompts.
+- **Onboarding a user**: \`mutagent login --browser --json\` — CLI prints auth URL to stdout, polls 5 min. **Surface the URL verbatim to the user.** \`--non-interactive\` is NOT needed when \`--browser\` is set.
+
+\`mutagent login\` is canonical. \`mutagent auth login\` is a back-compat alias. Both delegate to \`lib/auth-flow.ts\`. Decision record: [cli-design-principles.md](../../docs/cli-design-principles.md) → Login Unification.
+
+---
+
+## Extensibility
+
+Add \`workflows/custom-<name>.md\` with frontmatter \`triggers: ["phrase"]\` — auto-discovered by the decision tree fallback row. No rebuild needed.
+`,
+  "concepts/eval-criteria.md": `---
+name: mutagent-cli-concepts-eval-criteria
+description: |
+  Canonical source for MutagenT evaluation-criteria framing:
+  INPUT-param criteria → Minimum Viable Context (MVC);
+  OUTPUT-param criteria → Output Standards.
+  3-tier rubric discipline: grounded, observable, never vague.
+  Includes current platform validation rules for criterion shape.
+  Mirrored in mutagent/src/modules/prompts/prompt-evaluations/README.md.
+triggers:
+  - "evaluation criteria"
+  - "eval criteria"
+  - "rubric"
+  - "how do I evaluate"
+  - "minimum viable context"
+  - "output standards"
+  - "evaluationParameter"
+  - "guided eval"
+---
+
+# Concept — Guided Evaluation Criteria
+
+> **Canonical source** for the INPUT vs OUTPUT framing. Mirrored in:
+> - \`mutagent-cli/src/commands/prompts/evaluation/guided-workflow.ts\` — kept in sync.
+> - \`mutagent/src/modules/prompts/prompt-evaluations/README.md\` as BE-side
+>   reference — mirror, not a fork.
+
+## The golden rule
+
+**Every evaluation criterion is scoped to EITHER the inputs OR the outputs of
+the prompt. Never both in one criterion.**
+
+Mixing them produces vague criteria ("is the output good given the inputs?")
+that the optimizer cannot act on. When collecting rubrics from the user,
+explicitly frame each question around one of these two scopes.
+
+---
+
+## NEVER auto-generate criteria
+
+This is Rule 3 of the [5 Core Rules](../SKILL.md). Reasons:
+
+- Auto-generated rubrics are vague by default ("score based on accuracy")
+- The optimizer cannot act on vague criteria — it needs observable tiers
+- The user knows what "correct" means for their domain; the agent does not
+- Generic rubrics produce noisy scores that mislead the optimizer
+
+**Always use AskUserQuestion to collect rubrics, one per variable/dimension.**
+
+When \`evaluation create --guided --json\` runs, the CLI provides a list of
+fields that need rubrics. Ask the user about EVERY field in that list.
+Do not skip any field. Do not pre-fill answers. The user must provide each rubric.
+
+---
+
+## Input-param criteria → Minimum Viable Context (MVC)
+
+**Scope**: the \`{variables}\` the prompt template consumes. Each \`{variable}\` is
+an **input param**. The criterion asks: *is the information required for the
+prompt to succeed actually present?*
+
+### 3-tier completeness scale (grounded, never vague)
+
+- \`1.0\` — all required context present, no ambiguity
+- \`0.5\` — some context present, enough to attempt but likely partial/hedged answer
+- \`0.0\` — critical context missing, prompt cannot succeed regardless of model
+
+### Discipline rules
+
+- **Grounded**: every tier must describe an observable property of the input
+  data, not a feeling.
+- **Observable**: a human reader should assign a tier by looking at a single
+  input row — no model output needed.
+- **Per-variable**: one criterion per \`{variable}\`. This localizes optimizer signal.
+
+### Before asking the user
+
+Enumerate variables using the delimiter inferred by \`mutagent explore --json\`:
+- \`delimiter: "single"\` → \`{foo}\` — platform canonical
+- \`delimiter: "double"\` → \`{{foo}}\` — framework template; convert before upload
+
+See [concepts/prompt-variables.md](./prompt-variables.md) for the full inference contract.
+
+### Example
+
+For \`"Summarize {document} for {audience}"\`:
+
+\`\`\`json
+[
+  {
+    "name": "document-present",
+    "evaluationParameter": "document",
+    "description": "1.0 = actual prose text >= 100 chars. 0.5 = short snippet or summary stub. 0.0 = filename, null, TODO, or empty."
+  },
+  {
+    "name": "audience-concrete",
+    "evaluationParameter": "audience",
+    "description": "1.0 = concrete persona (e.g. 'junior Python devs'). 0.5 = broad category ('engineers'). 0.0 = empty, 'general', 'everyone'."
+  }
+]
+\`\`\`
+
+Note the 1:1 mapping: one criterion per variable, \`evaluationParameter\` equal
+to the variable name.
+
+---
+
+## Output-param criteria → Output Standards
+
+**Scope**: the model's response shape and content. Pick the dimensions relevant
+to the task and write one criterion per dimension.
+
+### Common dimensions
+
+- **Content correctness** — right answer, right facts, right tone
+- **Structural correctness** — matches \`outputSchema\`, required fields, enums, length
+- **Groundedness** — facts in the output traceable to facts in the input
+- **Format compliance** — JSON validity, markdown shape, regex match
+
+### 3-tier correctness scale
+
+Concrete pass/fail per tier. **Never** write "0-1 scale" or "score based on
+accuracy" — those are vague and the optimizer can't act on them.
+
+### Example (JSON-output summarizer)
+
+\`\`\`json
+{
+  "name": "summary-correctness",
+  "evaluationParameter": "summary",
+  "description": "1.0 = valid JSON, all 3 required fields, covers all key arguments. 0.5 = valid JSON, 1-2 arguments missing/hedged. 0.0 = invalid JSON OR fabricated facts."
+}
+\`\`\`
+
+The \`evaluationParameter\` here is the output-schema field name, not an input
+variable. Same 1:1 discipline — one criterion per output dimension.
+
+---
+
+## Platform validation rules (current)
+
+When creating an evaluation via \`mutagent prompts evaluation create -d '<json>'\`,
+each criterion must pass these platform-enforced checks:
+
+| Field | Required | Validation |
+|---|---|---|
+| \`name\` | yes | slug-like, no spaces, \`[a-z0-9-_]\` |
+| \`description\` | yes | non-empty, ideally >= 20 chars with tier definitions |
+| \`evaluationParameter\` | yes | must match a variable name from the prompt OR an output field name |
+
+**Common validation failures:**
+- \`evaluationParameter\` references a variable not in the prompt template → rejected
+- \`description\` is too short or vague → accepted by platform but produces poor scores
+- Multiple criteria with the same \`evaluationParameter\` → accepted but wasteful
+
+---
+
+## How to apply when creating an evaluation
+
+1. **Read the prompt template.** Enumerate \`{variables}\` (input params) +
+   expected output shape (from \`outputSchema\` or the code's parse logic).
+2. **Ask the user**: "Evaluate INPUTS (is context sufficient) or OUTPUTS
+   (is response correct) first?" — let the user pick the scope.
+3. **Collect criteria**: use AskUserQuestion to collect from user, never auto-generate — one per variable (INPUT) or per dimension (OUTPUT),
+   always with a 3-tier rubric describing observable behavior.
+4. **Map to platform shape**:
+   \`\`\`typescript
+   {
+     name: string;                // short, slug-like
+     description: string;         // the 3-tier rubric verbatim
+     evaluationParameter: string; // the variable name OR output field
+   }
+   \`\`\`
+5. **Upload** via \`mutagent prompts evaluation create <id> -d '<json>' --json\`.
+
+The \`--guided\` flag walks the user through this flow interactively — use it
+when the user is new to the concept. Follow the CLI's next-step guidance in
+the output to collect rubrics in the correct order.
+
+---
+
+## Anti-patterns
+
+- **Auto-generating criteria** — Rule 3: NEVER. Always collect from user.
+- **Mixing input and output in one criterion** — breaks signal; split into two.
+- **Vague rubrics** — "0.8 if mostly good" → rewrite with observable tiers.
+- **One criterion for many variables** — reduces signal, slows optimization.
+- **Scoring the model, not the data** — MVC scores the INPUT data quality.
+
+---
+
+## Cross-references
+
+- [SKILL.md](../SKILL.md) → 5 rules (Rule 3: never auto-generate)
+- [workflows/optimization.md](../workflows/optimization.md) → steps 7-9 (where this concept is applied)
+- [concepts/prompt-variables.md](./prompt-variables.md) → delimiter inference (used in MVC step)
+- \`mutagent/src/modules/prompts/prompt-evaluations/README.md\` → BE mirror
+`,
+  "concepts/prompt-variables.md": `---
+name: mutagent-cli-concepts-prompt-variables
+description: |
+  Prompt template variable delimiter inference contract.
+  Platform canonical is single-brace {variable}. Third-party frameworks vary
+  (Handlebars / Mustache / Liquid / Jinja2 use double {{variable}}).
+  \`mutagent explore\` infers the delimiter per file and surfaces it in
+  \`--json\` output as \`delimiter: "single" | "double"\`.
+  Includes conversion rules for upload and apply phases.
+triggers:
+  - "prompt variables"
+  - "template variables"
+  - "single vs double brace"
+  - "{variable}"
+  - "{{variable}}"
+  - "delimiter"
+  - "inferPromptVariables"
+  - "brace convention"
+  - "convert variables"
+---
+
+# Concept — Prompt Variables
+
+## Platform canonical
+
+**MutagenT platform uses single-brace \`{variable}\`.** The platform renders
+prompts by substituting \`{name}\` with the provided value at optimization /
+evaluation time.
+
+## Third-party framework variance
+
+Real-world codebases use different delimiters depending on which prompt
+framework the user already has installed:
+
+| Framework | Delimiter | Example |
+|---|---|---|
+| **MutagenT platform** (canonical) | single | \`{document}\` |
+| **LangChain** \`PromptTemplate\` | single | \`{document}\` |
+| **LangChain** \`ChatPromptTemplate\` + Mustache | double | \`{{document}}\` |
+| **Handlebars** | double | \`{{document}}\` |
+| **Mustache** | double | \`{{document}}\` |
+| **LiquidJS** | double | \`{{ document }}\` |
+| **Jinja2** (Python) | double | \`{{ document }}\` |
+
+---
+
+## Brace conversion — upload and apply
+
+Getting this wrong breaks templates after optimization. Follow the two-phase rule:
+
+### Phase 1 — Upload (code → MutagenT)
+
+If the code has \`{{double}}\` braces:
+1. Warn the user: "Your template uses \`{{double}}\` braces (Handlebars/LangChain Mustache). MutagenT uses \`{single}\` braces. I'll convert before uploading."
+2. Convert \`{{name}}\` → \`{name}\` in the prompt content passed to \`mutagent prompts create\`.
+3. Record the original delimiter in \`.mutagent/mutation-context.md\` so the apply phase knows to convert back.
+
+If the code has \`{single}\` braces: no conversion needed — upload as-is.
+
+### Phase 2 — Apply (MutagenT → code)
+
+After optimization, the platform returns a prompt with \`{single}\` braces.
+
+If the original codebase used \`{{double}}\` braces:
+1. Convert \`{name}\` → \`{{name}}\` in the optimized prompt before writing to the source file.
+2. Confirm with the user before saving.
+
+If the original codebase used \`{single}\` braces: write the optimized prompt as-is.
+
+**Summary table:**
+
+| Code uses | Upload | Optimized output | Write back to code |
+|---|---|---|---|
+| \`{single}\` | as-is | \`{single}\` | as-is |
+| \`{{double}}\` | convert to \`{single}\` | \`{single}\` | convert back to \`{{double}}\` |
+
+---
+
+## Per-file inference — \`mutagent explore\`
+
+The CLI's \`mutagent explore\` command calls \`inferPromptVariables()\` on every
+matching source file and **infers the delimiter per file** rather than
+globally. A single repository may contain both LangChain \`PromptTemplate\`
+(single) and Handlebars email templates (double) side by side.
+
+### Inference algorithm
+
+1. **Strip fenced markdown code blocks** first (\` \`\`\` ... \`\`\` \`). Avoids false
+   positives from prompts that document JSON examples in fenced blocks.
+2. **Count \`{{name}}\` matches** → \`doubleHits\`.
+3. **Count \`{name}\` matches** that are NOT adjacent to \`{\` (not part of \`{{...}}\`)
+   and NOT followed by \`"\` (JSON-key skip) → \`singleHits\`.
+4. **Majority wins**: \`doubleHits > singleHits\` → \`double\`, else \`single\`.
+5. **Tie-break**: singleHits === doubleHits (including the 0/0 case) →
+   \`single\` (platform canonical).
+
+### Escaped-JSON caveat
+
+Prompts like \`"Return {\\"status\\": \\"ok\\"}"\` — the \`{\` is followed by \`"\`, so
+the single-brace regex deliberately skips it. Never treat literal JSON keys as
+template variables.
+
+---
+
+## How to use the delimiter field
+
+\`mutagent explore --json\` surfaces the inferred delimiter per discovered prompt:
+
+\`\`\`json
+{
+  "prompts": [
+    {
+      "file": "src/prompts/summarize.ts",
+      "line": 12,
+      "preview": "const prompt = \`Summarize {document} for {audience}\`;",
+      "reason": "template-variable",
+      "confidence": "high",
+      "delimiter": "single"
+    },
+    {
+      "file": "src/emails/welcome.hbs",
+      "line": 3,
+      "preview": "<p>Hello {{name}}, welcome to {{product}}!</p>",
+      "reason": "template-variable",
+      "confidence": "high",
+      "delimiter": "double"
+    }
+  ]
+}
+\`\`\`
+
+Use the delimiter field to:
+- Enumerate variables correctly (don't treat \`{{foo}}\` as two \`{foo}\` tokens).
+- Decide whether to convert before upload (Phase 1 above).
+- Drive the [concepts/eval-criteria.md](./eval-criteria.md) → MVC step (one criterion per variable).
+
+---
+
+## Edge cases
+
+- **Empty prompt** — no variables → tie (0/0) → \`single\` (canonical).
+- **Mixed delimiters in one file** — majority wins, tied files default to \`single\`.
+  Warn the user: their codebase probably has two prompt systems co-existing.
+- **Nested braces** \`{{{ foo }}}\` — Handlebars triple-brace (no-escape). Currently
+  matched by the double regex as \`{{ foo }}\`; outer \`}\` ignored. Fine for inference.
+
+---
+
+## Cross-references
+
+- [SKILL.md](../SKILL.md) → 5 rules + journey router
+- [workflows/optimization.md](../workflows/optimization.md) → step 3 (delimiter drives variable enumeration) and step 15 (apply conversion)
+- [concepts/eval-criteria.md](./eval-criteria.md) → MVC (Minimum Viable Context) — uses delimiter to enumerate input params
+- Source: \`mutagent-cli/src/lib/explorer.ts\` → \`inferPromptVariables()\` and \`DiscoveredPrompt.delimiter\`
+- Tests: \`mutagent-cli/src/__tests__/lib/explorer.test.ts\`
+`,
+  "workflows/agents.md": `---
+name: mutagent-cli-workflows-agents
+description: |
+  Agent path (multi-turn, tool-calling) — WORK IN PROGRESS.
+  Agent optimization & evaluation are actively in development.
+  This workflow short-circuits to a partnership link, offers read-only CRUD,
+  and optionally extracts a sub-prompt back to the Optimization path.
+triggers:
+  - "agent"
+  - "langchain agent"
+  - "langgraph"
+  - "crewai"
+  - "autogen"
+  - "openai agents"
+  - "tool calling"
+  - "multi-turn"
+  - "AgentExecutor"
+  - "createReactAgent"
+  - "StateGraph"
+---
+
+# Workflow — Agents (WIP)
+
+> **This path is WORK IN PROGRESS.** Agent optimization and evaluation are
+> actively in development. Do NOT run a multi-turn agent through the prompt
+> optimizer — the platform will reject it, and scores would be meaningless for
+> a tool-calling loop.
+
+Read the **5 rules** in [SKILL.md](../SKILL.md) before executing.
+
+---
+
+## When this workflow applies
+
+Load this file when the journey router in [SKILL.md](../SKILL.md) matched one of these
+signals in the user's code:
+
+- \`AgentExecutor\`, \`createReactAgent\`, \`createToolCallingAgent\`, \`createStructuredChatAgent\`
+- \`from "openai/agents"\` or \`from "@openai/agents"\`
+- \`from "crewai"\`, \`from "autogen"\`, \`from "autogen_agentchat"\`
+- \`from "@langchain/langgraph"\`, \`StateGraph(\`
+- \`tool_calls\` / \`toolCalls\` property access
+- \`@tool\` decorator (Python)
+- \`tools: [...]\` array in an LLM call config
+- \`function_call\` / \`tool_choice\` fields
+- \`while\` loop + LLM call
+
+\`mutagent explore --json\` flags these under \`agents[]\` (not \`prompts[]\`).
+
+---
+
+## Required card (show this verbatim to the user)
+
+When the agent path is triggered, copy this card into your chat response
+verbatim — do NOT paraphrase, do NOT collapse into a bash block:
+
+\`\`\`
+I see you have an Agent (multi-turn / tool-calling). Agent Optimization &
+Evaluations are actively in development in MutagenT. For early access and
+to partner with us on the roadmap:
+→ https://www.mutagent.io/agents-partnership
+\`\`\`
+
+---
+
+## Sequence
+
+\`\`\`
+1. Run \`mutagent explore --json\` if you haven't already.
+   → Confirm agents[] is non-empty.
+   → Show command output to user.
+
+2. Show the WIP card above verbatim in your chat response.
+
+3. Use AskUserQuestion to explain that agent code cannot be directly optimized:
+   "Your code looks like a multi-turn agent. I can't run it
+   through the prompt optimizer yet. Would you like to:
+   (a) Join early access → https://www.mutagent.io/agents-partnership
+   (b) Inspect existing agents in MutagenT (read-only CRUD)
+   (c) Extract a single sub-prompt inside the agent loop and optimize it"
+
+4. Branch:
+   (a) → surface the URL verbatim. STOP.
+   (b) → \`mutagent agents list --json\` ; \`mutagent agents get <id> --json\`
+          → show results to user. STOP (no mutations available).
+   (c) → extract one sub-prompt, then route to [workflows/optimization.md](./optimization.md)
+          treating the sub-prompt as a standalone Prompt.
+\`\`\`
+
+---
+
+## Branch (c) — extracting a sub-prompt
+
+Multi-turn agents often contain inner prompts that ARE suitable for the
+Optimization path:
+
+- a planner prompt ("given this user goal, list the tools you'd call")
+- a summarizer prompt ("given these tool results, write the final answer")
+- a classifier prompt ("which tool should handle this input?")
+
+Each of these is a single-shot prompt with a clear output schema. Extract ONE,
+treat it as a standalone Prompt.
+
+When extracting:
+1. Identify the exact string literal or template that becomes the sub-prompt.
+2. Enumerate its \`{variables}\` per [concepts/prompt-variables.md](../concepts/prompt-variables.md).
+3. Confirm with user: "I'll optimize the planner prompt only, not the full agent. Sound right?"
+4. On confirmation → load [workflows/optimization.md](./optimization.md) from step 3 (prompts create).
+
+Do NOT try to extract the whole agent loop at once.
+
+---
+
+## What NOT to do
+
+- **Do not** call \`mutagent prompts optimize start\` on an agent file.
+- **Do not** upload an agent's full system prompt + tool definitions as a single "prompt".
+- **Do not** suggest "try it anyway" — the WIP status is deliberate.
+- **Do not** skip showing the WIP card to the user — they need the partnership URL.
+
+---
+
+## Cross-references
+
+- [SKILL.md](../SKILL.md) → 5 rules + journey router
+- [workflows/exploration.md](./exploration.md) → where \`agents[]\` entries are first detected
+- [workflows/optimization.md](./optimization.md) → branch (c) destination
+- [concepts/prompt-variables.md](../concepts/prompt-variables.md) → \`{foo}\` vs \`{{foo}}\` for sub-prompt extraction
+- Partnership link: https://www.mutagent.io/agents-partnership
+`,
+  "workflows/exploration.md": `---
+name: mutagent-cli-workflows-exploration
+description: |
+  Read-only codebase scan workflow. Discovers prompts and agents, classifies
+  each as optimization-eligible (prompt) or WIP (agent), and presents a
+  structured taxonomy to the user before any write operations.
+triggers:
+  - "explore"
+  - "scan codebase"
+  - "find prompts"
+  - "what prompts do I have"
+  - "discover prompts"
+  - "what's in my codebase"
+---
+
+# Workflow — Exploration (Read-Only Discovery)
+
+> **This workflow is read-only.** No writes, no uploads, no mutations.
+> Use it to understand what's in the codebase before deciding next steps.
+
+Read the **5 rules** in [SKILL.md](../SKILL.md) before executing. Key reminders:
+- \`--json\` on every command
+- \`<command> --help\` before first use
+- This workflow ends with a user question — never auto-proceed to writes
+
+---
+
+## When this workflow applies
+
+- User said "explore my codebase", "what prompts do I have", "find prompts", "scan"
+- Intent is unclear and you need to discover before acting
+- User wants to understand what's optimizable before committing to a path
+
+---
+
+## Prompt vs Agent — taxonomy
+
+Before running explore, understand what the CLI will return:
+
+**PROMPT (optimization-eligible)**
+- Single LLM call: input → LLM → output
+- Template with \`{variables}\` (single) or \`{{variables}}\` (double brace)
+- Can be evaluated with G-Eval (input/output pair)
+- Can be optimized by Metatuner
+- Appears in \`prompts[]\` in explore output
+
+**AGENT (NOT directly optimizable)**
+- Multi-turn loop: input → LLM → tool → LLM → ... → output
+- Dynamic branching execution
+- Cannot be optimized like a prompt (no fixed output schema)
+- Requires agent-level evaluation (not yet available)
+- Appears in \`agents[]\` in explore output
+
+### Agent detection heuristics (what \`mutagent explore\` flags)
+
+| Pattern | Framework | Classification |
+|---|---|---|
+| \`new StateGraph(\` / \`StateGraph.compile()\` | LangGraph | agent |
+| \`new Agent(\` / \`agent.execute()\` | Mastra / custom | agent |
+| \`AgentExecutor\` / \`createReactAgent\` / \`createToolCallingAgent\` | LangChain | agent |
+| \`from "openai/agents"\` / \`from "@openai/agents"\` | OpenAI Agents SDK | agent |
+| \`from "crewai"\` / \`from "autogen"\` | CrewAI / AutoGen | agent |
+| \`tools: [...]\` + LLM call in a loop | custom | agent |
+| \`while\` loop + LLM call | custom agent loop | agent |
+| \`@tool\` decorator (Python) | any | agent |
+| \`tool_calls\` / \`toolCalls\` property access | any | agent |
+| Single \`openai.chat()\` or template string | — | prompt |
+| String with \`{variable}\` / \`{{variable}}\` | — | prompt |
+
+---
+
+## Workflow steps
+
+\`\`\`
+1. mutagent explore --json
+   → surfaces: prompts[], agents[], datasets[], markers[]
+   → show command output to user
+
+2. Classify results:
+   - prompts[]  → optimization-eligible (single-shot, output schema)
+   - agents[]   → WIP (multi-turn/tool-calling) — route to [workflows/agents.md](./agents.md)
+   - datasets[] → existing local data (uploadable in optimization workflow)
+   - markers[]  → already-uploaded items (show dashboard links)
+
+3. Note the \`delimiter\` field on each prompt entry:
+   - "single" → {variable} — MutagenT native, no conversion needed
+   - "double" → {{variable}} — framework template, conversion required on upload
+   See [concepts/prompt-variables.md](../concepts/prompt-variables.md) for the conversion rules.
+
+4. Use AskUserQuestion to present findings and ask which prompts to upload:
+   "Here's what I found in your codebase:
+    - N prompt(s) found: [list files]
+    - N agent(s) found: [list files] (WIP — not optimizable yet)
+    - N dataset(s) found: [list]
+    - N already-uploaded: [list]
+    What would you like to do?"
+
+5. Route based on user answer:
+   - "optimize this prompt" → load [workflows/optimization.md](./optimization.md)
+   - "add tracing" → load [workflows/tracing.md](./tracing.md)
+   - "tell me about the agent" → load [workflows/agents.md](./agents.md)
+   - "nothing yet" → STOP (read-only complete)
+\`\`\`
+
+---
+
+## Output handling
+
+After step 1, show the command output to the user before proceeding to classification. Do NOT proceed to step 2 until the user has seen the results.
+
+---
+
+## Brace convention note
+
+\`mutagent explore --json\` surfaces \`delimiter: "single" | "double"\` per discovered prompt. Use this before deciding how to enumerate variables. See [concepts/prompt-variables.md](../concepts/prompt-variables.md) for the full inference contract and conversion rules.
+
+---
+
+## Common pitfalls
+
+- Skipping the classification step → user gets a raw JSON dump instead of a next-action recommendation
+- Treating \`agents[]\` entries as optimization-eligible → they are NOT; route to [workflows/agents.md](./agents.md)
+- Auto-proceeding to writes after explore → always confirm with user first
+- Ignoring the \`delimiter\` field → wrong variable enumeration when uploading a double-brace prompt
+
+---
+
+## Cross-references
+
+- [SKILL.md](../SKILL.md) → 5 rules + journey router
+- [concepts/prompt-variables.md](../concepts/prompt-variables.md) → \`{foo}\` vs \`{{foo}}\` inference + conversion
+- [workflows/optimization.md](./optimization.md) → next step after exploration (prompt path)
+- [workflows/tracing.md](./tracing.md) → next step after exploration (integration path)
+- [workflows/agents.md](./agents.md) → next step after exploration (agent path)
+`,
+  "workflows/optimization.md": `---
+name: mutagent-cli-workflows-optimization
+description: |
+  Full prompt optimization journey: explore → prompts create → dataset add →
+  evaluation create (guided) → optimize start → watch/status → results → apply.
+  Enforces 5 rules: --json always, --help before first use, user-collected
+  eval criteria, explore-before-modify, cost transparency before optimize.
+triggers:
+  - "optimize prompt"
+  - "improve prompt"
+  - "tune prompt"
+  - "evaluate prompt"
+  - "upload prompt"
+  - "create evaluation"
+  - "upload dataset"
+  - "run optimizer"
+  - "start optimization"
+---
+
+# Workflow — Optimization (Full Journey)
+
+> **This is the full loop.** Expect 5-10 CLI calls and at least one long-running
+> optimizer job. Each step requires user confirmation. Never auto-run the full
+> chain without presenting findings at each gate.
+
+Read the **5 rules** in [SKILL.md](../SKILL.md) before executing. All 5 rules apply here:
+- \`--json\` on every command (Rule 1)
+- \`--help\` before first use of any command (Rule 2)
+- **NEVER auto-generate eval criteria** — collect from user (Rule 3)
+- \`mutagent explore --json\` before any write (Rule 4)
+- \`mutagent usage --json\` before \`optimize start\` (Rule 5)
+
+---
+
+## When this workflow applies
+
+- User said "optimize prompt", "improve prompt", "tune prompt"
+- User wants to upload a prompt and measure its quality
+- User wants to run the Metatuner optimizer
+
+---
+
+## Required pre-reads (load these before the relevant steps)
+
+| Step | Pre-read | Why |
+|---|---|---|
+| Before \`prompts create\` | [concepts/prompt-variables.md](../concepts/prompt-variables.md) | Brace convention — single \`{var}\` vs double \`{{var}}\` affects how variables are parsed |
+| Before \`evaluation create --guided\` | [concepts/eval-criteria.md](../concepts/eval-criteria.md) | INPUT MVC + OUTPUT Standards — 3-tier rubric format |
+
+---
+
+## Directive chain
+
+\`\`\`
+explore → prompts create → dataset add → evaluation create --guided
+  → [Use AskUserQuestion to collect rubrics from the user for each field]
+  → evaluation create -d '<json>'
+  → usage check
+  → [Use AskUserQuestion to confirm optimization cost with user]
+  → optimize start
+  → optimize status (poll)
+  → optimize results
+  → [Use AskUserQuestion to present scorecard: Apply / View Diff / Reject]
+  → On Apply: Edit source file
+\`\`\`
+
+---
+
+## Full workflow steps
+
+\`\`\`
+ 1. mutagent explore --json
+    → find candidate prompts in codebase
+    → show command output to user
+    → confirm with user: "Which prompt would you like to optimize?"
+
+ 2. mutagent prompts --help
+    mutagent prompts create --help
+    → read flags before using (Rule 2)
+
+ 3. Load [concepts/prompt-variables.md](../concepts/prompt-variables.md)
+    → determine if prompt uses {single} or {{double}} braces
+    → if double-brace: warn user about conversion requirement
+
+ 4. mutagent prompts create --name <name> [--system-file / --raw-file] --json
+    → show command output to user
+    → record promptId from response
+
+ 5. mutagent prompts dataset add --help
+    → read flags before using
+
+ 6. mutagent prompts dataset add <promptId> -d '[...]' --name "<name>" --json
+    → upload dataset rows (input/output pairs)
+    → show command output to user
+    → record datasetId
+
+ 7. Load [concepts/eval-criteria.md](../concepts/eval-criteria.md)
+    → understand INPUT-param (MVC) vs OUTPUT-param (Standards) scope
+
+ 8. mutagent prompts evaluation create <promptId> --guided --json
+    → the CLI provides a list of fields, each needing a rubric
+    → follow the CLI's next-step guidance in the output
+    → for EVERY field listed (INPUT scope first, then OUTPUT):
+        - ask the user the provided question for that field
+        - wait for user response
+        - do NOT skip any field
+        - do NOT auto-generate any answer
+    → collect at minimum: one INPUT criterion per {variable}, one OUTPUT criterion
+
+ 9. mutagent prompts evaluation create <promptId> -d '<json>' --json
+    → upload the criteria collected in step 8
+    → show command output to user
+    → record evaluationId
+
+10. mutagent usage --json
+    → show usage/quota to user
+    → confirm with user: "This optimization will use N iterations (~X min each).
+       You have Y remaining. Proceed?"
+    → STOP if user declines
+
+11. mutagent prompts optimize start <promptId> \\
+      --dataset <datasetId> \\
+      --evaluation <evaluationId> \\
+      --max-iterations 1 \\
+      --json
+    → NEVER set --max-iterations > 1 without explicit user consent
+    → record jobId from response
+
+12. mutagent prompts optimize status <jobId> --json
+    → poll until status = "completed" or "failed"
+    → show progress to user
+
+13. mutagent prompts optimize results <jobId> --json
+    → ALWAYS show before/after scorecard to user
+
+14. Confirm with user: "Here's the before/after scorecard. What would you like to do?
+    (a) Apply — update the prompt in your source file
+    (b) View diff first
+    (c) Reject — keep the original"
+
+15. On Apply (a): Edit the prompt in the user's source file
+    → replace old prompt text with optimized version
+    → if double-brace codebase: convert {variable} back to {{variable}}
+    → confirm with user before saving
+\`\`\`
+
+---
+
+## Cost control
+
+- Default \`--max-iterations 1\` is the only value you may use without explicit consent.
+- If user requests more: confirm the number with user → confirm the cost implication.
+- Each iteration = one full G-Eval run over the dataset × LLM calls. This costs real money.
+
+---
+
+## Apply / Reject rules
+
+- **Apply**: edit the source file with the optimized prompt. If the codebase used \`{{double}}\` braces, convert the optimized \`{single}\` brace output back to \`{{double}}\` before writing. See [concepts/prompt-variables.md](../concepts/prompt-variables.md) → Conversion.
+- **Reject**: no file changes. Record the jobId in \`.mutagent/mutation-context.md\` for future reference.
+- **View diff**: show a unified diff of old vs new prompt text before asking again.
+
+---
+
+## Common pitfalls
+
+- Running \`optimize start\` before \`evaluation create\` → optimizer has no scoring signal
+- Mixing INPUT and OUTPUT criteria in the same rubric → vague scores
+- Applying results without showing the before/after scorecard first
+- Forgetting to convert \`{single}\` back to \`{{double}}\` after apply in double-brace codebases
+- Starting with \`--max-iterations 3\` without consent
+
+---
+
+## Guided Dataset Creation
+
+When no local dataset exists, use the guided mode to curate high-quality test data:
+
+\`\`\`
+mutagent prompts dataset add <prompt-id> --guided --json
+\`\`\`
+
+The CLI analyzes the prompt's inputSchema + outputSchema and returns:
+- **Suggested categories**: edge cases, hard cases, representative cases
+- **Per-field questions**: what values, what edge cases, what correct output looks like
+- **Template item**: showing the expected shape for each dataset entry
+- **Priority rule**: hard cases that expose prompt weaknesses > easy cases that always pass
+
+Collect answers from the user, then construct 5-10 dataset items covering all categories.
+Ensure at least 2 hard/edge cases per category. Then upload:
+
+\`\`\`
+mutagent prompts dataset add <prompt-id> -d '<constructed-json>' --name '<name>' --json
+\`\`\`
+
+---
+
+## Cross-references
+
+- [SKILL.md](../SKILL.md) → 5 rules + journey router
+- [concepts/prompt-variables.md](../concepts/prompt-variables.md) → brace convention + conversion (critical for steps 3 and 15)
+- [concepts/eval-criteria.md](../concepts/eval-criteria.md) → INPUT MVC + OUTPUT Standards + 3-tier rubric (critical for steps 7-8)
+- [workflows/exploration.md](./exploration.md) → step 1 of this workflow
+- [workflows/tracing.md](./tracing.md) → parallel or follow-up path
+`,
+  "workflows/tracing.md": `---
+name: mutagent-cli-workflows-tracing
+description: |
+  Framework integration workflow. Adds MutagenT tracing/observability to the
+  user's codebase. Non-destructive (append-only), fastest first-value path.
+  Detects framework via \`mutagent explore\`, generates snippet via
+  \`mutagent integrate\`, applies via Edit tool, verifies via traces list.
+triggers:
+  - "add tracing"
+  - "add observability"
+  - "integrate framework"
+  - "integrate mutagent"
+  - "add mutagent to my code"
+  - "instrument my prompts"
+  - "trace my prompts"
+---
+
+# Workflow — Tracing (Framework Integration)
+
+> **Scope**: read + append-only on user code. Never modify existing business
+> logic — only add tracing imports and decorators.
+
+Read the **5 rules** in [SKILL.md](../SKILL.md) before executing. Key reminders:
+- \`--json\` on every command
+- \`<command> --help\` before first use
+- Explore before modify (Rule 4)
+- Show command output to user after every mutation
+
+---
+
+## When this workflow applies
+
+- User said "add tracing", "add observability", "integrate \\<framework\\>"
+- User wants to see their prompts captured in the MutagenT dashboard
+- Fastest path to first value — prefer this before suggesting optimization
+
+---
+
+## Supported frameworks
+
+\`mutagent integrate --help\` is the authoritative list. Common entries:
+
+| Framework | Signal in codebase |
+|---|---|
+| LangChain | \`from "langchain"\`, \`PromptTemplate\`, \`ChatPromptTemplate\` |
+| LangGraph | \`from "@langchain/langgraph"\`, \`StateGraph\` |
+| OpenAI SDK | \`import OpenAI\`, \`openai.chat.completions.create\` |
+| Vercel AI SDK | \`import { generateText }\` from \`"ai"\` |
+| Mastra | \`from "@mastra/core"\` |
+| Custom / raw | any string template with \`{variable}\` |
+
+---
+
+## Workflow steps
+
+\`\`\`
+1. mutagent explore --json
+   → detect which framework is in use
+   → show command output to user
+
+2. mutagent integrate --help
+   → read the available frameworks and flags (ALWAYS before step 3)
+
+3. mutagent integrate <framework> --json
+   → get the integration snippet
+
+4. Use AskUserQuestion to confirm before applying code changes: "I'll add the tracing snippet to <file>. Proceed?"
+   → show the snippet preview before applying
+
+5. Apply snippet via Edit tool
+   → code change happens here — append imports + decorators only
+   → never touch existing business logic
+
+6. mutagent traces list --json
+   → verify integration: check traces count > 0
+   → show results and dashboard link to user
+\`\`\`
+
+---
+
+## Scope guard
+
+This path is **READ + APPEND-ONLY** on the user's code:
+- ✓ Add import at top of file
+- ✓ Wrap existing function call with tracing decorator
+- ✗ Rename variables
+- ✗ Refactor logic
+- ✗ Remove existing code
+
+If the integration snippet requires a significant rewrite, confirm scope with the user before proceeding.
+
+---
+
+## Post-integration state
+
+After step 6:
+- Update \`.mutagent/mutation-context.md\` with the integration marker
+- Show the dashboard link from traces output so user can verify traces in UI
+- If user wants to optimize the traced prompt → route to [workflows/optimization.md](./optimization.md)
+
+---
+
+## Common pitfalls
+
+- Skipping step 1 and guessing the framework → let \`explore\` detect it
+- Forgetting step 6 → user has no proof the integration works
+- Editing files outside the \`integrate --json\` snippet block
+- Not showing command results to the user after mutations
+
+---
+
+## Cross-references
+
+- [SKILL.md](../SKILL.md) → 5 rules + journey router
+- [workflows/exploration.md](./exploration.md) → step 1 of this workflow
+- [workflows/optimization.md](./optimization.md) → natural next step after tracing
+- [concepts/prompt-variables.md](../concepts/prompt-variables.md) → variable inference for traced prompts
+`
+};
+var SKILL_CONTENT = SKILL_FILES["SKILL.md"] ?? "";
+
+// src/commands/skills.ts
+function findRepoRoot() {
+  try {
+    return execSync4("git rev-parse --show-toplevel", {
+      encoding: "utf-8",
+      stdio: ["pipe", "pipe", "pipe"]
+    }).trim();
+  } catch {
+    return process.cwd();
+  }
+}
+var SKILL_DIR = ".claude/skills/mutagent-cli";
+var SKILL_ENTRY_FILE = "SKILL.md";
+function assertSkillFilesValid() {
+  if (Object.keys(SKILL_FILES).length === 0) {
+    throw new Error("SKILL_FILES is missing or empty. This indicates a broken build: " + "scripts/sync-skill.ts must regenerate src/generated/skill-content.ts " + "from .claude/skills/mutagent-cli/ before build/test. " + "Run `bun run sync-skill` to debug.");
+  }
+  const entry = SKILL_FILES[SKILL_ENTRY_FILE];
+  if (typeof entry !== "string" || entry.trim().length === 0) {
+    throw new Error(`SKILL_FILES["${SKILL_ENTRY_FILE}"] is missing or empty. ` + "Check .claude/skills/mutagent-cli/SKILL.md and run `bun run sync-skill`.");
+  }
+  if (!entry.startsWith("---")) {
+    throw new Error(`SKILL_FILES["${SKILL_ENTRY_FILE}"] is malformed: expected YAML frontmatter ("---") at start.`);
+  }
+}
+function getSkillFiles() {
+  assertSkillFilesValid();
+  return SKILL_FILES;
+}
+function createSkillsCommand() {
+  const skills = new Command17("skills").description("Manage MutagenT CLI skills for coding agents");
+  skills.command("install").description("Install MutagenT CLI skill for Claude Code").addHelpText("after", `
+Examples:
+  ${chalk36.dim("$")} mutagent skills install
+
+This creates a Claude Code skill at .claude/skills/mutagent-cli/SKILL.md
+that teaches coding agents how to use the MutagenT CLI effectively.
+    `).action((_options, cmd) => {
+    const parentCmd = cmd.parent?.parent;
+    const isJson = parentCmd ? getJsonFlag(parentCmd) : false;
+    const output = new OutputFormatter(isJson ? "json" : "table");
+    const repoRoot = findRepoRoot();
+    const skillDir = join7(repoRoot, SKILL_DIR);
+    const files = getSkillFiles();
+    const writtenFiles = [];
+    let totalBytes = 0;
+    const sortedKeys = Object.keys(files).sort((a, b) => {
+      if (a === SKILL_ENTRY_FILE)
+        return -1;
+      if (b === SKILL_ENTRY_FILE)
+        return 1;
+      return a.localeCompare(b);
+    });
+    for (const relPath of sortedKeys) {
+      const destPath = join7(skillDir, relPath);
+      const parentDir = dirname(destPath);
+      if (!existsSync12(parentDir)) {
+        mkdirSync4(parentDir, { recursive: true });
+      }
+      const raw = files[relPath] ?? "";
+      const finalContent = raw.endsWith(`
+`) ? raw : `${raw}
+`;
+      writeFileSync5(destPath, finalContent, "utf-8");
+      writtenFiles.push({ path: destPath, bytes: finalContent.length });
+      totalBytes += finalContent.length;
+    }
+    if (isJson) {
+      output.output({
+        installed: true,
+        name: "mutagent-cli",
+        skillDir,
+        files: writtenFiles,
+        totalBytes
+      });
+    } else {
+      output.success(`Installed MutagenT CLI skill`);
+      console.log(`  ${chalk36.dim("Dir:")} ${skillDir}`);
+      console.log("");
+      for (let i = 0;i < writtenFiles.length; i++) {
+        const entry = writtenFiles[i];
+        if (!entry)
+          continue;
+        const relPath = sortedKeys[i] ?? "";
+        const isLast = i === writtenFiles.length - 1;
+        const prefix = isLast ? "└─" : "├─";
+        console.log(`  ${chalk36.dim(prefix)} ${relPath} ${chalk36.dim(`(${String(entry.bytes)} bytes)`)}`);
+      }
+      console.log("");
+      console.log(`  ${chalk36.dim(`${String(sortedKeys.length)} file(s), ${String(totalBytes)} bytes total.`)}`);
+      console.log(`  ${chalk36.dim("This skill teaches coding agents how to use the MutagenT CLI.")}`);
+      console.log(`  ${chalk36.dim("It will be automatically loaded by Claude Code when relevant triggers match.")}`);
+    }
+  });
+  return skills;
+}
+
+// src/commands/usage.ts
+init_config();
+import { Command as Command18 } from "commander";
+import chalk37 from "chalk";
+init_errors();
+init_sdk_client();
+var PROVIDERS_URL = "https://app.mutagent.io/settings/providers";
+function createUsageCommand() {
+  const usage = new Command18("usage").description("Show resource counts (prompts, datasets, evaluations, optimizations, experiments)").addHelpText("after", `
+Examples:
+  ${chalk37.dim("$")} mutagent usage
+  ${chalk37.dim("$")} mutagent usage --json
+    `);
+  usage.action(async () => {
+    const isJson = getJsonFlag(usage);
     const output = new OutputFormatter(isJson ? "json" : "table");
     try {
       const apiKey = getApiKey();
@@ -9143,17 +10544,17 @@ Examples:
         });
       } else {
         console.log("");
-        console.log(chalk36.bold("\uD83D\uDCCA MutagenT Usage"));
-        console.log(chalk36.dim("─".repeat(45)));
+        console.log(chalk37.bold("\uD83D\uDCCA MutagenT Usage"));
+        console.log(chalk37.dim("─".repeat(45)));
         console.log("");
-        console.log(chalk36.bold("Resources:"));
-        console.log(`  Prompts:         ${chalk36.cyan(String(promptCount))}`);
-        console.log(`  Datasets:        ${chalk36.cyan(String(datasetCount))}`);
-        console.log(`  Evaluations:     ${chalk36.cyan(String(evaluationCount))}`);
-        console.log(`  Optimizations:   ${chalk36.cyan(String(optimizationCount))}`);
-        console.log(`  Experiments:     ${chalk36.cyan(String(experimentCount))}`);
+        console.log(chalk37.bold("Resources:"));
+        console.log(`  Prompts:         ${chalk37.cyan(String(promptCount))}`);
+        console.log(`  Datasets:        ${chalk37.cyan(String(datasetCount))}`);
+        console.log(`  Evaluations:     ${chalk37.cyan(String(evaluationCount))}`);
+        console.log(`  Optimizations:   ${chalk37.cyan(String(optimizationCount))}`);
+        console.log(`  Experiments:     ${chalk37.cyan(String(experimentCount))}`);
         console.log("");
-        console.log(`  Providers: ${chalk36.underline(PROVIDERS_URL)}`);
+        console.log(`  Providers: ${chalk37.underline(PROVIDERS_URL)}`);
         console.log("");
       }
     } catch (error) {
@@ -9441,11 +10842,11 @@ Claude Code Session Telemetry:
 
 // src/commands/feedback.ts
 import { Command as Command20 } from "commander";
-import chalk37 from "chalk";
+import chalk38 from "chalk";
 init_errors();
 init_config();
 import { readFileSync as readFileSync11 } from "fs";
-import { join as join9, dirname } from "path";
+import { join as join9, dirname as dirname2 } from "path";
 import { fileURLToPath } from "url";
 var VALID_CATEGORIES = ["bug", "feature", "improvement", "praise"];
 function getCliVersion() {
@@ -9453,7 +10854,7 @@ function getCliVersion() {
     return process.env.CLI_VERSION;
   }
   try {
-    const __dirname2 = dirname(fileURLToPath(import.meta.url));
+    const __dirname2 = dirname2(fileURLToPath(import.meta.url));
     const pkgPath = join9(__dirname2, "..", "..", "package.json");
     const pkg = JSON.parse(readFileSync11(pkgPath, "utf-8"));
     return pkg.version ?? "0.1.1";
@@ -9495,12 +10896,12 @@ async function postToServer(payload, endpoint, apiKey, workspaceId, organization
 }
 function createFeedbackCommand() {
   const feedback = new Command20("feedback").description("Send product feedback to MutagenT").addHelpText("after", `
-${chalk37.bold("Examples:")}
-  ${chalk37.cyan('mutagent feedback send -m "Great optimization results!"')}
-  ${chalk37.cyan('mutagent feedback send -m "CLI crashed on export" --category bug')}
-  ${chalk37.cyan('mutagent feedback send -m "Need batch operations" --category feature --json')}
+${chalk38.bold("Examples:")}
+  ${chalk38.cyan('mutagent feedback send -m "Great optimization results!"')}
+  ${chalk38.cyan('mutagent feedback send -m "CLI crashed on export" --category bug')}
+  ${chalk38.cyan('mutagent feedback send -m "Need batch operations" --category feature --json')}
 
-${chalk37.yellow("AI Agent (MANDATORY):")}
+${chalk38.yellow("AI Agent (MANDATORY):")}
   ALWAYS use --json: mutagent feedback send -m "..." --category improvement --json
   Use this command to report bugs, request features, or share UX feedback.
     `).action(() => {
@@ -9511,18 +10912,18 @@ ${chalk37.yellow("AI Agent (MANDATORY):")}
 }
 function registerFeedbackSend(feedback) {
   feedback.command("send").description("Send feedback about the MutagenT platform").requiredOption("-m, --message <text>", "Feedback message").option("--category <type>", `Feedback category: ${VALID_CATEGORIES.join(", ")}`, "improvement").option("--session <id>", "Link feedback to a specific session").addHelpText("after", `
-${chalk37.bold("Examples:")}
-  ${chalk37.dim("$")} mutagent feedback send -m "The optimization UX could show progress better"
-  ${chalk37.dim("$")} mutagent feedback send -m "CLI errored on traces export" --category bug
-  ${chalk37.dim("$")} mutagent feedback send -m "Love the guided eval!" --category praise --json
-
-${chalk37.bold("Categories:")}
-  ${chalk37.bold("bug")}          Something is broken or not working as expected
-  ${chalk37.bold("feature")}      Request a new capability
-  ${chalk37.bold("improvement")}  Suggest a UX or workflow enhancement (default)
-  ${chalk37.bold("praise")}       Share what you love about the platform
-
-${chalk37.yellow("AI Agent (MANDATORY):")}
+${chalk38.bold("Examples:")}
+  ${chalk38.dim("$")} mutagent feedback send -m "The optimization UX could show progress better"
+  ${chalk38.dim("$")} mutagent feedback send -m "CLI errored on traces export" --category bug
+  ${chalk38.dim("$")} mutagent feedback send -m "Love the guided eval!" --category praise --json
+
+${chalk38.bold("Categories:")}
+  ${chalk38.bold("bug")}          Something is broken or not working as expected
+  ${chalk38.bold("feature")}      Request a new capability
+  ${chalk38.bold("improvement")}  Suggest a UX or workflow enhancement (default)
+  ${chalk38.bold("praise")}       Share what you love about the platform
+
+${chalk38.yellow("AI Agent (MANDATORY):")}
   ALWAYS use --json: mutagent feedback send -m "..." --json
   Auto-captured context (CLI version, platform, node version) is included automatically.
     `).action(async (options) => {
@@ -9579,7 +10980,7 @@ if (process.env.CLI_VERSION) {
   cliVersion = process.env.CLI_VERSION;
 } else {
   try {
-    const __dirname2 = dirname2(fileURLToPath2(import.meta.url));
+    const __dirname2 = dirname3(fileURLToPath2(import.meta.url));
     const pkgPath = join10(__dirname2, "..", "..", "package.json");
     const pkg = JSON.parse(readFileSync12(pkgPath, "utf-8"));
     cliVersion = pkg.version ?? cliVersion;
@@ -9594,90 +10995,100 @@ program.name("mutagent").description(`MutagenT CLI - AI-native prompt optimizati
   showGlobalOptions: true
 });
 program.addHelpText("after", `
-${chalk38.yellow("Non-Interactive Mode (CI/CD & Coding Agents):")}
-  export MUTAGENT_API_KEY=mt_...  ${chalk38.dim("or")}  --api-key mt_...
-  --json ${chalk38.dim("for structured output")}    --non-interactive ${chalk38.dim("to disable prompts")}
-
-${chalk38.yellow("Command Navigation:")}
-  mutagent login                        ${chalk38.dim("Login (browser OAuth — recommended)")}
-  mutagent auth status                  ${chalk38.dim("Check auth + workspace")}
-  mutagent init                         ${chalk38.dim("Initialize project (.mutagentrc.json)")}
-  mutagent explore                      ${chalk38.dim("Discover prompts in codebase")}
-  mutagent workspaces list --json       ${chalk38.dim("List workspaces (verify ID)")}
-  mutagent config set workspace <id>    ${chalk38.dim("Set active workspace")}
-  mutagent usage --json                 ${chalk38.dim("Show account usage + provider status")}
-
-  mutagent prompts create --help        ${chalk38.dim("Upload prompt (read help first!)")}
-  mutagent prompts list --json          ${chalk38.dim("List prompts")}
-  mutagent prompts get <id> --json      ${chalk38.dim("Full prompt details + schemas")}
-
-  mutagent prompts dataset add --help   ${chalk38.dim("Upload dataset (read help first!)")}
-  mutagent prompts dataset list <id>    ${chalk38.dim("List datasets")}
-
-  mutagent prompts evaluation create --help           ${chalk38.dim("Create eval (read help first!)")}
-  mutagent prompts evaluation create <id> --guided --json   ${chalk38.dim("Guided eval workflow")}
-  mutagent prompts evaluation list <id> --json        ${chalk38.dim("List evaluations")}
-
-  mutagent prompts optimize start --help      ${chalk38.dim("Run optimization (read help first!)")}
-  mutagent prompts optimize status <job-id>   ${chalk38.dim("Poll progress")}
-  mutagent prompts optimize results <job-id>  ${chalk38.dim("View scorecard")}
-
-  mutagent feedback send -m "..."         ${chalk38.dim("Send product feedback")}
-  mutagent feedback send -m "..." --category bug  ${chalk38.dim("Report a bug")}
-
-  mutagent integrate <framework>        ${chalk38.dim("Framework integration guide")}
-  mutagent hooks --help                 ${chalk38.dim("Hook setup for Claude Code telemetry")}
-  mutagent playground run <id> --input '{...}'  ${chalk38.dim("Quick test")}
-
-${chalk38.yellow("★ Workflow: Framework Integration (Tracing):")}
-  1. mutagent explore                                    ${chalk38.dim("← discover prompts/agents in codebase")}
-  2. mutagent integrate <framework>                      ${chalk38.dim("← get integration instructions")}
-  3. Apply tracing code to your codebase                 ${chalk38.dim("← follow the guide output")}
-  4. mutagent traces list --json                         ${chalk38.dim("← verify traces are arriving")}
-
-${chalk38.yellow("★ Workflow: Evaluate → Optimize:")}
-  1. mutagent prompts create --help               ${chalk38.dim("← read help")}
-  2. mutagent prompts create ... --json            ${chalk38.dim("← upload prompt with {variables} + inputSchema")}
-  3. mutagent prompts dataset add --help           ${chalk38.dim("← read help")}
-  4. mutagent prompts dataset add <id> ... --json  ${chalk38.dim("← upload dataset")}
-  5. mutagent prompts evaluation create <id> --guided --json  ${chalk38.dim("← guided eval")}
+${AGENTS_WIP_BANNER}
+
+${chalk39.yellow("Non-Interactive Mode (CI/CD & Coding Agents):")}
+  export MUTAGENT_API_KEY=mt_...  ${chalk39.dim("or")}  --api-key mt_...
+  --json ${chalk39.dim("for structured output")}    --non-interactive ${chalk39.dim("to disable prompts")}
+
+${chalk39.yellow("Command Navigation:")}
+  mutagent login                        ${chalk39.dim("Login (browser OAuth — recommended)")}
+  mutagent auth status                  ${chalk39.dim("Check auth + workspace")}
+  mutagent init                         ${chalk39.dim("Initialize project (.mutagentrc.json)")}
+  mutagent explore                      ${chalk39.dim("Discover prompts in codebase")}
+  mutagent workspaces list --json       ${chalk39.dim("List workspaces (verify ID)")}
+  mutagent config set workspace <id>    ${chalk39.dim("Set active workspace")}
+  mutagent usage --json                 ${chalk39.dim("Show account usage + provider status")}
+
+  mutagent prompts create --help        ${chalk39.dim("Upload prompt (read help first!)")}
+  mutagent prompts list --json          ${chalk39.dim("List prompts")}
+  mutagent prompts get <id> --json      ${chalk39.dim("Full prompt details + schemas")}
+
+  mutagent prompts dataset add --help   ${chalk39.dim("Upload dataset (read help first!)")}
+  mutagent prompts dataset list <id>    ${chalk39.dim("List datasets")}
+
+  mutagent prompts evaluation create --help           ${chalk39.dim("Create eval (read help first!)")}
+  mutagent prompts evaluation create <id> --guided --json   ${chalk39.dim("Guided eval workflow")}
+  mutagent prompts evaluation list <id> --json        ${chalk39.dim("List evaluations")}
+
+  mutagent prompts optimize start --help      ${chalk39.dim("Run optimization (read help first!)")}
+  mutagent prompts optimize status <job-id>   ${chalk39.dim("Poll progress")}
+  mutagent prompts optimize results <job-id>  ${chalk39.dim("View scorecard")}
+
+  mutagent feedback send -m "..."         ${chalk39.dim("Send product feedback")}
+  mutagent feedback send -m "..." --category bug  ${chalk39.dim("Report a bug")}
+
+  mutagent integrate <framework>        ${chalk39.dim("Framework integration guide")}
+  mutagent hooks --help                 ${chalk39.dim("Hook setup for Claude Code telemetry")}
+  mutagent playground run <id> --input '{...}'  ${chalk39.dim("Quick test")}
+
+${chalk39.yellow("★ Workflow: Framework Integration (Tracing):")}
+  1. mutagent explore                                    ${chalk39.dim("← discover prompts/agents in codebase")}
+  2. mutagent integrate <framework>                      ${chalk39.dim("← get integration instructions")}
+  3. Apply tracing code to your codebase                 ${chalk39.dim("← follow the guide output")}
+  4. mutagent traces list --json                         ${chalk39.dim("← verify traces are arriving")}
+
+${chalk39.yellow("★ Workflow: Evaluate → Optimize:")}
+  1. mutagent prompts create --help               ${chalk39.dim("← read help")}
+  2. mutagent prompts create ... --json            ${chalk39.dim("← upload prompt with {variables} + inputSchema")}
+  3. mutagent prompts dataset add --help           ${chalk39.dim("← read help")}
+  4. mutagent prompts dataset add <id> ... --json  ${chalk39.dim("← upload dataset")}
+  5. mutagent prompts evaluation create <id> --guided --json  ${chalk39.dim("← guided eval")}
   6. mutagent prompts optimize start <id> --dataset <d> --evaluation <e> --json
 
-${chalk38.yellow("Post-Onboarding Decision Tree:")}
-  After ${chalk38.bold("mutagent auth login")}, users land in one of 3 paths:
-  ${chalk38.bold("Path A")} (Tracing):      explore → integrate <framework> → apply tracing → verify
-  ${chalk38.bold("Path B")} (Optimization): explore → prompts create → dataset add → eval create → optimize
-  ${chalk38.bold("Path C")} (Manual):       Use CLI commands directly — run mutagent <command> --help
+${chalk39.yellow("Post-Onboarding Decision Tree:")}
+  After ${chalk39.bold("mutagent auth login")}, users land in one of 3 paths:
+  ${chalk39.bold("Path A")} (Tracing):      explore → integrate <framework> → apply tracing → verify
+  ${chalk39.bold("Path B")} (Optimization): explore → prompts create → dataset add → eval create → optimize
+  ${chalk39.bold("Path C")} (Manual):       Use CLI commands directly — run mutagent <command> --help
 
-${chalk38.yellow("Directive System:")}
+${chalk39.yellow("Directive System:")}
   Every --json response may include:
-  ${chalk38.bold("_directive.renderedCard")}  Pre-formatted card for the user ${chalk38.red("(MUST be shown in chat)")}
-  ${chalk38.bold("_directive.instruction")}   Next step for the agent
-  ${chalk38.bold("_directive.next")}          Array of suggested follow-up commands
-  ${chalk38.bold("_links")}                   Dashboard/API URLs (format as markdown links)
+  ${chalk39.bold("_directive.renderedCard")}  Pre-formatted card for the user ${chalk39.red("(MUST be shown in chat)")}
+  ${chalk39.bold("_directive.instruction")}   Next step for the agent
+  ${chalk39.bold("_directive.next")}          Array of suggested follow-up commands
+  ${chalk39.bold("_links")}                   Dashboard/API URLs (format as markdown links)
 
-${chalk38.yellow("Evaluation Criteria Format:")}
-  Each criterion MUST have: ${chalk38.bold("name")}, ${chalk38.bold("description")} (scoring rubric), ${chalk38.bold("evaluationParameter")}
+${chalk39.yellow("Evaluation Criteria Format:")}
+  Each criterion MUST have: ${chalk39.bold("name")}, ${chalk39.bold("description")} (scoring rubric), ${chalk39.bold("evaluationParameter")}
   evaluationParameter MUST match an inputSchema or outputSchema field name
   No duplicate evaluationParameter values — each criterion targets a unique field
   ALL schema fields must be covered (missing fields = error)
-  Use ${chalk38.bold("--guided --json")} to generate criteria templates from prompt schemas
-
-${chalk38.yellow("Optimization Cost Control:")}
-  Default max-iterations is 1. ${chalk38.red("NEVER increase without explicit user request.")}
+  Use ${chalk39.bold("--guided --json")} to generate criteria templates from prompt schemas
+
+  ${chalk39.bold("Two framings — coding agents MUST distinguish:")}
+  ${chalk39.bold("• INPUT fields → Minimum Viable Context")}
+    What info MUST be present in the input for the prompt to succeed.
+    Rubric should include a ${chalk39.bold("completeness scale")} (e.g., 0 = missing critical context, 5 = complete).
+    ${chalk39.dim("Input fields are typically `{variables}` in your prompt template.")}
+  ${chalk39.bold("• OUTPUT fields → Output Standards")}
+    Content correctness + structural correctness + a ${chalk39.bold("correctness scale")} (e.g., 0 = wrong, 5 = perfect).
+
+${chalk39.yellow("Optimization Cost Control:")}
+  Default max-iterations is 1. ${chalk39.red("NEVER increase without explicit user request.")}
   Each iteration incurs LLM costs — confirm with user before starting >1.
 
-${chalk38.yellow("Post-Optimization:")}
-  After ${chalk38.bold("optimize results")}: ALWAYS show the before/after diff to the user first.
-  Then offer choices: ${chalk38.bold("Apply")} / ${chalk38.bold("Reject")}.
+${chalk39.yellow("Post-Optimization:")}
+  After ${chalk39.bold("optimize results")}: ALWAYS show the before/after diff to the user first.
+  Then offer choices: ${chalk39.bold("Apply")} / ${chalk39.bold("Reject")}.
 
-${chalk38.yellow("State Tracking:")}
+${chalk39.yellow("State Tracking:")}
   .mutagent/mutation-context.md — Codebase index of discovered/uploaded prompts
   Update after explore, create, and dataset operations
   mutagent auth status — Auth + workspace state
   Comment markers (// MutagenT:START ... // MutagenT:END) in source files
 
-${chalk38.yellow("AI Agent Rules (MANDATORY for coding agents):")}
+${chalk39.yellow("AI Agent Rules (MANDATORY for coding agents):")}
   1. Login (two paths):
      - CI / fully automated: export MUTAGENT_API_KEY=mt_... then mutagent login --json
      - Helping a user onboard: mutagent login --browser --json
@@ -9688,15 +11099,15 @@ ${chalk38.yellow("AI Agent Rules (MANDATORY for coding agents):")}
   3. Run <command> --help BEFORE first use of any command
   4. Use --guided --json for evaluation creation (NEVER --guided alone)
   5. Parse _directive.renderedCard and copy it into your CHAT RESPONSE verbatim
-     ${chalk38.red("HARD STOP")}: do NOT run further commands until the card is rendered in chat
+     ${chalk39.red("HARD STOP")}: do NOT run further commands until the card is rendered in chat
   6. After mutagent init, verify workspace: mutagent workspaces list --json
   7. Use {single_braces} for template variables in prompts
   8. Collect evaluation criteria from the user — NEVER auto-generate
   9. ALL user interaction via AskUserQuestion — CLI is non-interactive
 ${!hasCredentials() ? `
-` + chalk38.yellow("  Warning: Not authenticated. Run: mutagent login") + `
+` + chalk39.yellow("  Warning: Not authenticated. Run: mutagent login") + `
 ` : ""}${!hasRcConfig() ? `
-` + chalk38.green("  Get started: mutagent init") + `
+` + chalk39.green("  Get started: mutagent init") + `
 ` : ""}`);
 var rawArgs = process.argv.slice(2);
 if (rawArgs.includes("-v") || rawArgs.includes("--version")) {
@@ -9737,5 +11148,5 @@ program.addCommand(createHooksCommand());
 program.addCommand(createFeedbackCommand());
 program.parse();
 
-//# debugId=C3949CD3A9488F4264756E2164756E21
+//# debugId=4EA6EB9078B51CC264756E2164756E21
 //# sourceMappingURL=cli.js.map